A low-voltage network abnormality diagnosis method based on multi-source heterogeneous data fusion

Abstract

The application discloses a low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion, relates to the technical field of low-voltage network anomaly diagnosis and operation and maintenance, and comprises the following steps: collecting multi-source data, pre-processing the data by edge computing, adjusting the collection interval according to changes, processing the data by deep learning and other technologies, converting according to semantics, improving formula standardization, fusing the data by federated learning, dynamically weighting and intelligently interpolating, constructing a rule base, optimizing by reinforcement learning, setting a dynamic threshold, training a model based on Transform and GNN, improving the effect by contrast learning, monitoring data in real time, warning in multiple ways when an anomaly occurs, assisting in troubleshooting by AR, and repairing the anomaly by a robot. The application can fuse multi-source heterogeneous data and improve data quality, the diagnosis model and the rule base are accurate and efficient, anomalies can be found in time, the operation and maintenance cost is reduced through intelligent early warning, AR-assisted troubleshooting and robot repair, the stable operation of the low-voltage network is ensured, and the economic and social benefits are remarkable.

Description

Technical Field

[0001] This invention relates to the field of low-voltage power grid anomaly diagnosis and operation and maintenance technology, and in particular to a low-voltage power grid anomaly diagnosis method based on multi-source heterogeneous data fusion. Background Technology

[0002] With rapid economic development and ever-increasing electricity demand, the low-voltage power grid, as the end point of the power system, directly connects to a wide range of users. Its operational stability and reliability are crucial to ensuring normal electricity supply for these users. However, the low-voltage power grid is characterized by its wide distribution, complex structure, and diverse equipment types, making it prone to various abnormal situations during operation, such as voltage anomalies, current imbalances, and equipment failures.

[0003] Currently, traditional methods for diagnosing low-voltage network anomalies primarily rely on a single data source, such as analyzing data from distribution automation systems or electricity consumption information collection systems. This approach suffers from data bias, failing to comprehensively reflect the actual operating status of the low-voltage network. Furthermore, differences in data formats, standards, and meanings across different data sources make effective data integration and utilization difficult, increasing the complexity of anomaly diagnosis. In addition, traditional methods often employ fixed rules and models for anomaly identification, making them ill-suited to the complex and ever-changing operating environment of low-voltage networks and the emergence of new anomaly types, resulting in low accuracy and timeliness in anomaly diagnosis.

[0004] On the other hand, with the rapid development of smart grids, distributed energy, and the Internet of Things (IoT) technologies, a large number of distributed power sources, energy storage devices, and smart terminals have been connected to low-voltage networks, making the data sources of low-voltage networks more extensive and complex, exhibiting multi-source heterogeneous characteristics. How to effectively integrate these multi-source heterogeneous data, mine the potential information behind the data, and achieve accurate and timely diagnosis of low-voltage network anomalies has become an urgent problem to be solved in the power industry. Therefore, developing a low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion has significant practical implications. Summary of the Invention

[0005] This invention proposes a low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for diagnosing anomalies in low-voltage networks by fusing multi-source heterogeneous data includes the following steps: Data acquisition steps: Build a multi-source data acquisition platform, use edge computing devices to perform preliminary data processing at the source end, and adopt a time-series adaptive acquisition strategy to adjust the acquisition interval according to data changes; Data preprocessing steps: Use a deep learning-based outlier detection network to identify and correct outliers; use knowledge graph technology to classify and transform data based on semantic relationships. Data fusion steps: Each data source trains its model locally and then uploads the parameters to the fusion center. The fusion center uses an adaptive weighted fusion algorithm. w i To integrate weights, q i To score data quality, s i The importance of data sources is scored, and missing data is intelligently interpolated using a deep learning interpolation network. Rule base construction steps: Combine power industry standards, real-time monitoring data, and deep reinforcement learning algorithms to build a verification rule base, and use complex network analysis technology to mine hidden relationships in the data and transform them into rules; Model training steps: Construct a hybrid anomaly diagnosis model based on Transformer and graph neural networks, and introduce a contrastive loss function during training. , z i For feature vectors, Let be its positive samples, and sim be the similarity function; Real-time monitoring steps: A monitoring system is built based on a distributed stream computing framework. Sliding window technology is used for data aggregation and feature extraction. The processed data is then input into an anomaly diagnosis model and a verification rule base for parallel judgment. Early warning steps: When an anomaly is detected, alerts will be sent via SMS, pop-up windows, and intelligent voice alerts. The location of the anomaly will be displayed on an electronic map using GIS technology, and an automatic graded alert will be issued according to the severity of the anomaly. Troubleshooting assistance steps: Provide maintenance personnel with troubleshooting assistance tools based on augmented reality (AR) technology, combined with fault tree analysis and expert systems to guide troubleshooting; Anomaly repair steps: Use intelligent repair robots and remote control technology to repair anomalies, and then use detection equipment mounted on a drone to remotely check the results after repair.

[0007] Furthermore, it also includes a data quality assessment step. In the entire process of data collection, preprocessing, and fusion, a data quality assessment system based on blockchain and machine learning is built. The blockchain is used to record the data source, processing process, and quality assessment results. The machine learning algorithm is used to evaluate data quality from multiple dimensions. When the data quality assessment result is lower than the set threshold, the source of the data problem is automatically traced, and the data processing process is optimized and adjusted through smart contracts.

[0008] Furthermore, it also includes a model update step, which uses a combination of online learning and meta-learning techniques to dynamically update the anomaly diagnosis model, automatically adjust model parameters based on the characteristics of new data, evaluate the model regularly, and trigger the model update process when the model performance index drops below a certain threshold.

[0009] Furthermore, in the data acquisition step, multi-sensor equipment carried by drones is used to collect data from the low-voltage network. The drones conduct periodic inspections according to a preset flight path, and the collected data is transmitted to the acquisition platform in real time via a 5G network. By combining satellite remote sensing data and drone-collected data, information on the low-voltage network operating environment is obtained.

[0010] Furthermore, in the data preprocessing step, data standardization employs an improved normalization formula. , The mean, Standard deviation, The adjustment coefficient is used in this step, which employs a joint denoising algorithm based on wavelet transform and variational mode decomposition (VMD). According to the characteristics of noise at different scales, thresholding is performed in the wavelet domain. Finally, the data is reconstructed through inverse wavelet transform and inverse VMD transform. The thresholding process combines soft and hard thresholding, as shown in the formula. and ,in, λ represents the wavelet coefficients, and λ is the threshold.

[0011] Furthermore, in the data fusion step, quantum encryption technology is introduced to ensure the security of data fusion. Distributed data fusion technology based on blockchain and homomorphic encryption is adopted to perform fusion calculations on encrypted data. The homomorphic encryption algorithm supports specific calculations on ciphertext, thus achieving data fusion while ensuring data security.

[0012] Furthermore, in the rule base construction step, knowledge graphs and ontology reasoning techniques are used to mine and optimize the rules, construct a knowledge graph for the low-voltage network domain, associate and represent power equipment, operation data, and industry standard information, and use an ontology reasoning engine to perform reasoning on the knowledge graph to discover potential rules and abnormal patterns.

[0013] Furthermore, in the model training step, a combination of transfer learning and ensemble learning is adopted. Data is collected in different low-voltage network regions or at different time stages to train multiple basic models. Using transfer learning technology, the model parameters trained in one region or stage are transferred to models in other regions or stages for fine-tuning. Then, an ensemble learning algorithm is used to fuse the fine-tuned models, and the prediction results are obtained through voting or weighted averaging.

[0014] Furthermore, in the real-time monitoring step, an architecture combining edge computing and cloud computing is adopted. Edge computing devices are deployed on-site in the low-voltage network to perform real-time preprocessing and preliminary analysis of the collected data. The processed key data is then uploaded to the cloud computing platform for in-depth analysis and anomaly diagnosis. Data transmission protocols are used between the edge computing devices and the cloud computing platform, and load balancing technology is used to distribute computing tasks.

[0015] Furthermore, in the anomaly repair step, an anomaly repair knowledge base is established to record the process, methods, results, and related lessons learned from each anomaly repair. Natural language processing technology is used to extract and analyze the text information in the knowledge base. When a new anomaly is encountered, semantic matching is used to find similar anomaly cases in the knowledge base to provide reference and guidance for anomaly repair. At the same time, reinforcement learning algorithms are used to optimize the anomaly repair strategy.

[0016] Compared with existing technologies, the beneficial effects of this invention are: In terms of data processing, this method effectively integrates heterogeneous data from multiple sources. Through an advanced data acquisition platform, it not only covers traditional State Grid power supply system data but also integrates new data sources such as distributed energy monitoring and user-side smart home energy consumption, broadening the data sources. Simultaneously, by employing preprocessing techniques based on deep learning and semantic understanding, it can accurately identify and correct outliers, achieving reasonable data transformation and standardization, greatly improving data quality and usability, and providing a comprehensive and accurate data foundation for subsequent anomaly diagnosis.

[0017] Regarding anomaly diagnosis capabilities, this application constructs a hybrid model based on Transformer and graph neural networks, combined with contrastive learning techniques, which can fully capture the time-series features and topological relationships of data, effectively distinguishing different data features. Simultaneously, by incorporating a rule base based on deep reinforcement learning and complex network analysis, a dynamic threshold mechanism is introduced, which can dynamically adjust diagnostic rules according to different operating environments and conditions, greatly improving the accuracy and timeliness of anomaly diagnosis and enabling the rapid detection of various potential anomalies.

[0018] In practical application, this method provides a comprehensive anomaly handling solution. In the early warning stage, intelligent voice assistants, GIS technology, and a tiered early warning mechanism enable maintenance personnel to quickly understand anomalies. The troubleshooting assistance step utilizes AR technology and expert systems to provide maintenance personnel with intuitive and personalized troubleshooting paths. The anomaly repair step combines intelligent repair robots and remote control technology to achieve efficient and accurate anomaly repair. Furthermore, this method establishes a robust data quality assessment and model update mechanism, ensuring the system maintains consistently good performance, improving the stability and reliability of low-voltage network operation, and reducing maintenance costs, resulting in significant economic and social benefits. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of a low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion proposed in this invention; Figure 2 A bar chart comparing the accuracy of different methods in abnormality diagnosis; Figure 3 Line charts comparing response times for anomalies of varying complexity; Figure 4 A bar chart comparing operation and maintenance costs in different regions. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 4 A method for diagnosing low-voltage network anomalies through multi-source heterogeneous data fusion, comprising the following steps: S1: Data Acquisition Steps: Establish a highly compatible multi-source data acquisition platform. Obtain routine information such as equipment operation data and power consumption data from the State Grid power supply's electricity information acquisition system and distribution automation system; simultaneously connect to the distributed energy monitoring system to collect power generation data from distributed power sources, such as real-time power and power generation of photovoltaic power plants; and connect to user-side smart home energy consumption data to obtain the power consumption of household appliances. Deploy edge computing devices at the data acquisition site. These devices use low-power, high-performance chips, such as NVIDIA's Jetson Nano. The edge computing devices perform preliminary processing on the collected data, aggregating data from multiple sensors and extracting key features, such as calculating the average power over a period of time. Adjust the acquisition interval based on the frequency of data changes. Utilize time series analysis algorithms to analyze historical data and determine the change patterns of different data. For power data, if it fluctuates frequently over a period of time, shorten the acquisition interval to 5 minutes; for slowly changing data such as equipment archive information, extend the acquisition interval to 1 hour. This ensures the real-time nature of key data while reducing unnecessary data transmission and storage pressure.

[0024] S2: Data Preprocessing Steps: A deep learning-based outlier detection network is used, consisting of a Generative Adversarial Network (GAN) and a Convolutional Autoencoder (CAE). The generator attempts to generate samples similar to real data, the discriminator distinguishes between real and generated data, and the CAE reconstructs the data. When the reconstruction error exceeds a certain threshold, the data is identified as an outlier. For example, for voltage data, if the reconstruction error is greater than 5%, it is marked as an outlier and subsequently corrected through algorithms or manual intervention. Semantic understanding-based data transformation is achieved using knowledge graph technology. A knowledge graph for the low-voltage network domain is constructed, classifying data from different data sources according to their semantic relationships. For example, current data from different devices are categorized under the "current" node and further subdivided according to device type. This method allows for the reasonable transformation of categorical data. For data standardization, an improved normalization formula is used. ,in The mean of the data. Standard deviation, This is the adjustment coefficient (determined based on the data distribution, ranging from 0 to 0.5). For example, for power data with large fluctuations, Take 0.3; for relatively stable voltage data, Let's take 0.1. Data standardized using this formula better reflects the inherent characteristics of the data, improving the accuracy of subsequent analysis.

[0025] S3: Data Fusion Steps: Utilizing a federated learning-based data fusion architecture, data fusion is achieved without compromising the privacy of each data source. Each data source trains its model locally, uploading only the model parameters to the fusion center. The fusion center employs an adaptive weighted fusion algorithm, dynamically adjusting the fusion weights based on the quality and importance of the data from each data source. The formula is as follows: , where w i Let q be the fusion weight of the i-th data source. i To score data quality, s i The importance of the data source is scored. For missing data, intelligent interpolation is performed using an interpolation network based on Long Short-Term Memory (LSTM). Time series data is input into the LSTM network, which learns the temporal characteristics of the data and predicts missing values. For example, for missing current data at a certain moment, the LSTM network interpolates based on the current values ​​and trends of the preceding and following moments, ensuring data integrity.

[0026] S4: Rule Base Construction Steps: A verification rule base is constructed by combining power industry standards, real-time monitoring data, and deep reinforcement learning algorithms. Reinforcement learning algorithms (such as the Proximal Policy Optimization (PPO) algorithm) interact with the environment (the actual operation of the low-voltage network) to continuously explore and optimize rules. A dynamic threshold mechanism is introduced to dynamically adjust voltage and current thresholds based on factors such as season and electricity consumption periods. For example, during peak summer electricity consumption periods, the voltage threshold can be appropriately relaxed by 5% from the normal range. Simultaneously, complex network analysis techniques are used to treat devices and data in the low-voltage network as nodes and edges, constructing a complex network. The connection relationships and data flow of nodes in the network are analyzed to uncover potential anomaly patterns, such as how an anomaly in one device might trigger a chain reaction in connected devices. These patterns are then transformed into rules and added to the rule base.

[0027] S5: Model Training Steps: Construct a hybrid anomaly diagnostic model based on Transformer and Graph Neural Network (GNN). Transformer is used to capture the time-series features of the data, while GNN is used to analyze the topological relationships between data points. Contrastive learning techniques are employed to enhance the model's ability to distinguish different data features, and a contrastive loss function is introduced during training. , where z i For data feature vectors, Using positive samples as the model and sim as the similarity function, the model parameters are optimized through backpropagation to improve the model's generalization ability and anomaly detection accuracy.

[0028] S6: Real-time Monitoring Steps: A real-time monitoring system is built based on a distributed stream computing framework (such as Apache Flink) to process and analyze low-voltage network data in real time. A sliding window technique is employed, setting different window sizes and sliding steps according to data characteristics. For example, for voltage data, the window size is set to 1 minute and the sliding step to 10 seconds. Data aggregation and feature extraction are performed within the window. The processed data is simultaneously input into an anomaly diagnosis model and a verification rule base for parallel judgment, improving the real-time performance and accuracy of monitoring.

[0029] S7: Warning Procedure: When an anomaly is detected, an alert is sent to maintenance personnel via SMS, pop-up window, and intelligent voice assistant. The SMS message details the time and location of the anomaly (accurate to the transformer substation or line name), the type of anomaly (e.g., overvoltage, current imbalance), and the potential impact range. The pop-up window is displayed in a prominent red on the maintenance personnel's monitoring terminal to attract their attention. The intelligent voice assistant broadcasts the anomaly information via voice, ensuring timely awareness even during busy periods. Utilizing Geographic Information System (GIS) technology, the anomaly location is visually displayed on an electronic map, with different colored icons indicating the anomaly type and shading to represent the impact range. Based on the severity of the anomaly, alerts are categorized into different levels: severe anomalies are marked in red, and general anomalies in yellow, allowing maintenance personnel to quickly assess the urgency and take appropriate measures.

[0030] S8: Troubleshooting Assistance Steps: Equip maintenance personnel with AR glasses based on augmented reality (AR) technology. Upon arrival at the site, maintenance personnel scan the equipment using the AR glasses. A virtual model of the equipment will be displayed on the glasses, labeled with its parameters such as rated voltage and rated current, as well as historical anomaly records. Combining fault tree analysis and an expert system, a visualized troubleshooting path is generated based on the anomaly type. For example, when a voltage anomaly occurs, the fault tree starts from the voltage anomaly node, with branches showing possible causes such as transformer failure or line short circuit. The expert system provides troubleshooting suggestions based on the actual site conditions, guiding maintenance personnel to quickly locate the cause of the fault and improve troubleshooting efficiency.

[0031] S9: Anomaly Repair Steps: For remotely operable devices, such as smart switches and smart meters, repairs are performed via remote control commands. Remote control software is used to send control signals to the device, adjust device parameters, or execute corresponding operations. For example, when a smart meter experiences a communication failure, a restart command is sent to attempt to restore communication. For devices requiring on-site operation, an intelligent repair robot is dispatched. The intelligent repair robot uses a Deep Q-Network (DQN) algorithm for path planning and operation decisions. Based on on-site environmental information and anomaly information, the robot plans a path to the faulty device and selects appropriate repair actions according to the DQN algorithm. After repair, a drone equipped with detection equipment is used to remotely detect the repair effect, such as checking whether voltage and current have returned to normal, to ensure successful repair.

[0032] This invention also includes a data quality assessment step. A data quality assessment system based on blockchain and machine learning is constructed throughout the entire process of data acquisition, preprocessing, and fusion. The immutability of blockchain is used to record the source of data, the processing procedure, and the quality assessment results. Machine learning algorithms (such as random forest algorithms) are used to assess data quality from multiple dimensions. Assessment indicators include not only data integrity, accuracy, and consistency, but also timeliness and reliability. When the data quality assessment result falls below a set threshold, the source of the data problem is automatically traced, and the data processing flow is optimized and adjusted through smart contracts to ensure data quality.

[0033] This invention also includes a model update step. A combination of online learning and meta-learning techniques is used to dynamically update the anomaly diagnosis model. Online learning algorithms (such as the Adaptive Stochastic Gradient Descent algorithm ASGD) process new data in real time and continuously adjust model parameters. Meta-learning algorithms (such as the gradient-based meta-learning algorithm MAML) learn how to quickly adapt to new tasks and data, automatically adjusting hyperparameters such as the learning rate and network structure based on the characteristics of the new data. The model is periodically and comprehensively evaluated. When model performance metrics (such as accuracy, recall, and F1 score) drop beyond a certain threshold (such as 5%), the model update process is triggered to ensure the model maintains consistently good performance.

[0034] In this invention, during the data acquisition step, the drone is equipped with multiple high-performance sensors, including a high-resolution optical camera capable of clearly capturing the appearance of low-voltage power grid lines, accurately detecting even minor wear and corrosion; an infrared thermal imager effectively senses the temperature distribution of the lines, promptly detecting temperature anomalies caused by overload or poor contact; and a lidar system is used to perform three-dimensional mapping of the geographical environment of the low-voltage network, accurately delineating the terrain and the spatial relationships of surrounding objects. The drone conducts regular inspections according to a flight path set by a professional path planning algorithm. During flight, it travels stably at a predetermined speed and altitude, ensuring the comprehensiveness and consistency of the collected data. The acquired data is transmitted to the acquisition platform in real time using the high-speed transmission capabilities of the 5G network. Furthermore, this invention combines satellite remote sensing data with drone-collected data. Satellite remote sensing can obtain large-scale geographical information and environmental data from a macroscopic perspective; the fusion of these two technologies provides richer and more complete data information for low-voltage network anomaly diagnosis, significantly improving the accuracy and reliability of the diagnosis.

[0035] In this invention, a joint denoising algorithm based on wavelet transform and variational mode decomposition (VMD) is employed in the data preprocessing step to address noise interference in the data. First, VMD is used to decompose the data into multiple intrinsic mode functions (IMFs). Then, wavelet transform is performed on each IMF. Based on the characteristics of noise at different scales, thresholding is performed in the wavelet domain. Finally, the data is reconstructed through inverse wavelet transform and inverse VMD transform. The thresholding process combines soft and hard thresholding, as shown in the formula: (Soft threshold portion) and (Hard thresholding) The thresholding method is adaptively selected based on the frequency characteristics and noise intensity of the IMF, effectively removing noise and improving data quality.

[0036] In this invention, quantum encryption technology is introduced to ensure the security of the data fusion process during the data fusion step. Quantum key distribution (QKD) technology is used to generate encryption keys, which are then used to encrypt the transmitted and fused data. Simultaneously, a distributed data fusion technology based on blockchain and homomorphic encryption is employed to perform fusion calculations on the encrypted data, ensuring that data privacy is not compromised. The homomorphic encryption algorithm supports specific calculations on the ciphertext, such as additive homomorphic encryption. Homomorphism with multiplication (E is the encryption function, m1 and m2 are plaintext data, and n and m are relevant parameters), achieving efficient data fusion while ensuring data security.

[0037] In this invention, knowledge graphs and ontology reasoning techniques are used for in-depth rule mining and optimization during the rule base construction step. A knowledge graph for the low-voltage network domain is constructed, representing information such as power equipment, operational data, and industry standards in a correlated manner. An ontology reasoning engine (such as the Pellet reasoning engine) is used to perform reasoning on the knowledge graph to discover potential rules and anomaly patterns. For example, based on the connection relationships between devices and operational data, possible fault propagation paths are inferred, transformed into rules, and added to the rule base, improving the intelligence and accuracy of the rule base.

[0038] In this invention, to further improve model performance during the model training step, a combination of transfer learning and ensemble learning is employed. Data is collected in different low-voltage network regions or at different time stages to train multiple base models. Using transfer learning techniques, the model parameters trained in one region or stage are transferred to models in other regions or stages for fine-tuning. Then, an ensemble learning algorithm (such as the random forest ensemble algorithm) is used to fuse the multiple fine-tuned models, and the final prediction result is obtained through voting or weighted averaging, enhancing the model's stability and generalization ability.

[0039] In this invention, an innovative architecture tightly integrating edge computing and cloud computing is employed in the real-time monitoring step. Edge computing devices equipped with high-performance processors and large-capacity storage are deployed at the low-voltage network site. These devices incorporate advanced data filtering algorithms that can quickly filter out noisy data and redundant information based on preset rules; they also utilize efficient feature extraction algorithms, such as deep learning-based convolutional neural network algorithms, to accurately extract key features such as voltage, current, and power. The key data, after preprocessing and preliminary analysis by the edge computing devices, is transmitted using the MQTT protocol. The MQTT protocol, with its lightweight design and publish / subscribe messaging model, can achieve reliable data transmission with extremely low bandwidth consumption in unstable network environments, greatly reducing latency. After the data is uploaded to the cloud computing platform, it is used for in-depth analysis using distributed computing frameworks such as Apache Hadoop and Spark. Simultaneously, load balancing technology is used to dynamically allocate computing tasks based on the resource usage of each computing node, avoiding single-point overload. Even under high load conditions, the system can ensure stable operation, significantly improving the efficiency and reliability of real-time monitoring and enabling timely and accurate detection of abnormal conditions in the low-voltage network.

[0040] In this invention, an anomaly repair knowledge base is established during the anomaly repair process. After each anomaly repair, the entire repair process is recorded in detail, including the specific context of the fault, the repair tools and equipment used, the details of each operation, the final repair result, and lessons learned. The advanced BERT model in natural language processing is used to deeply process the text information in the knowledge base. The BERT model, with its bidirectional Transformer architecture, can comprehensively understand the semantics of the text. Key information is accurately extracted through lexical analysis, syntactic analysis, and semantic role labeling. When a new anomaly occurs, similar cases are quickly searched in the knowledge base based on semantic vector similarity calculation, providing strategic references for the current repair. Simultaneously, reinforcement learning algorithms are introduced to optimize the anomaly repair strategy. Using repair time and repair effectiveness as reward mechanisms, the agent continuously explores different repair actions and states. Based on the feedback of the repair results, the strategy parameters are dynamically adjusted and continuously iterated and optimized, thereby significantly improving the success rate and efficiency of anomaly repair and ensuring the stable operation of the low-voltage network.

[0041] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for diagnosing low-voltage network anomalies through multi-source heterogeneous data fusion, characterized in that, Includes the following steps: Data acquisition steps: Build a multi-source data acquisition platform, use edge computing devices to perform preliminary data processing at the source end, and adopt a time-series adaptive acquisition strategy to adjust the acquisition interval according to data changes; Data preprocessing steps: Use a deep learning-based outlier detection network to identify and correct outliers; Using knowledge graph technology, data is classified and transformed based on semantic relationships. Data fusion steps: Each data source trains its model locally and then uploads the parameters to the fusion center. The fusion center uses an adaptive weighted fusion algorithm. w i To integrate weights, q i To score data quality, s i The importance of data sources is scored, and missing data is intelligently interpolated using a deep learning interpolation network. Rule base construction steps: Combine power industry standards, real-time monitoring data, and deep reinforcement learning algorithms to build a verification rule base, and use complex network analysis technology to mine hidden relationships in the data and transform them into rules; Model training steps: Construct a hybrid anomaly diagnosis model based on Transformer and graph neural networks, and introduce a contrastive loss function during training. , z i For feature vectors, Let be its positive samples, and sim be the similarity function; Real-time monitoring steps: A monitoring system is built based on a distributed stream computing framework. Sliding window technology is used for data aggregation and feature extraction. The processed data is then input into an anomaly diagnosis model and a verification rule base for parallel judgment. Early warning steps: When an anomaly is detected, alerts will be sent via SMS, pop-up windows, and intelligent voice alerts. The location of the anomaly will be displayed on an electronic map using GIS technology, and an automatic graded alert will be issued according to the severity of the anomaly. Troubleshooting assistance steps: Provide maintenance personnel with troubleshooting assistance tools based on augmented reality (AR) technology, combined with fault tree analysis and expert systems to guide troubleshooting; Anomaly repair steps: Use intelligent repair robots and remote control technology to repair anomalies, and then use detection equipment mounted on a drone to remotely check the results after repair.

2. The low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, It also includes a data quality assessment step. In the entire process of data collection, preprocessing, and fusion, a data quality assessment system based on blockchain and machine learning is built. The blockchain is used to record the data source, processing process and quality assessment results. The machine learning algorithm is used to evaluate the data quality from multiple dimensions. When the data quality assessment result is lower than the set threshold, the source of the data problem is automatically traced and the data processing process is optimized and adjusted through smart contracts.

3. The low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, It also includes a model update step, which uses a combination of online learning and meta-learning techniques to dynamically update the anomaly diagnosis model, automatically adjusts model parameters based on the characteristics of new data, evaluates the model regularly, and triggers the model update process when the model performance index drops below a certain threshold.

4. The low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In the data acquisition step, multi-sensor equipment carried by drones is used to collect data from the low-voltage network. The drones conduct periodic inspections according to a preset flight path, and the collected data is transmitted to the acquisition platform in real time via a 5G network. By combining satellite remote sensing data and drone-collected data, information on the low-voltage network operating environment is obtained.

5. The low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In the data preprocessing step, data standardization employs a modified normalization formula. , The mean, Standard deviation, The adjustment coefficient is used in this step, which employs a joint denoising algorithm based on wavelet transform and variational mode decomposition (VMD). According to the characteristics of noise at different scales, thresholding is performed in the wavelet domain. Finally, the data is reconstructed through inverse wavelet transform and inverse VMD transform. The thresholding process combines soft and hard thresholding, as shown in the formula. and ,in, λ represents the wavelet coefficients, and λ is the threshold.

6. The low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In the data fusion step, quantum encryption technology is introduced to ensure the security of data fusion. Distributed data fusion technology based on blockchain and homomorphic encryption is adopted to perform fusion calculations on encrypted data. Homomorphic encryption algorithm supports specific calculations on ciphertext, so as to achieve data fusion under the premise of ensuring data security.

7. The low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In the rule base construction step, knowledge graphs and ontology reasoning techniques are used to mine and optimize the rules, construct a knowledge graph for the low-voltage network domain, and associate and represent power equipment, operation data, and industry standard information. The ontology reasoning engine then performs reasoning on the knowledge graph to discover potential rules and abnormal patterns.

8. The low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In the model training step, a combination of transfer learning and ensemble learning is adopted. Data is collected in different low-voltage network regions or at different time stages to train multiple basic models. Using transfer learning technology, the model parameters trained in one region or stage are transferred to models in other regions or stages for fine-tuning. Then, an ensemble learning algorithm is used to fuse the fine-tuned models, and the prediction results are obtained through voting or weighted averaging.

9. The low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In the real-time monitoring step, an architecture combining edge computing and cloud computing is adopted. Edge computing devices are deployed on-site in the low-voltage network to perform real-time preprocessing and preliminary analysis of the collected data. The processed key data is then uploaded to the cloud computing platform for in-depth analysis and anomaly diagnosis. Data transmission protocols are used between the edge computing devices and the cloud computing platform, and load balancing technology is used to distribute computing tasks.

10. The low-voltage network anomaly diagnosis method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In the anomaly repair process, an anomaly repair knowledge base is established to record the process, methods, results, and related lessons learned from each anomaly repair. Natural language processing technology is used to extract and analyze the text information in the knowledge base. When a new anomaly is encountered, semantic matching is used to find similar anomaly cases in the knowledge base to provide reference and guidance for anomaly repair. At the same time, reinforcement learning algorithms are used to optimize the anomaly repair strategy.

Images