Power transformation primary equipment fault early warning method, system and equipment based on artificial intelligence

Abstract

The invention belongs to the technical field of power equipment fault early warning, and particularly relates to a power transformation primary equipment fault early warning method, system and device based on artificial intelligence. Each edge computing node uploads the trained model parameter updating information and the equipment health degree correction coefficient to a central server, and the central server updates the equipment health degree correction coefficient based on the received model parameter updating information of each edge computing node. A global fault early warning model is aggregated and generated in combination with a self-adaptive weight distribution mechanism including equipment topology correlation degree, a computing power fluctuation factor and an equipment health degree correction coefficient, and knowledge migration from a global model to a local model is completed by introducing a knowledge distillation technology of cross-equipment interference coupling degree. And realizing graded early warning by using the updated local fault early warning model. According to the method, the data privacy can be strictly protected, the differentiated characteristics of multiple power stations can be effectively adapted, and accurate fault early warning is realized.

Description

Technical Field

[0001] This invention belongs to the field of power equipment fault early warning technology, specifically relating to a method, system, equipment and medium for early warning of faults in primary substation equipment based on artificial intelligence. Background Technology

[0002] With the large-scale development of primary substation equipment, the impact of fault early warning technology on the stability of energy supply is becoming increasingly significant. Currently, mainstream fault early warning solutions in the industry generally adopt a centralized data processing model. This involves each substation uploading sensitive information such as equipment operating parameters and environmental monitoring data to a central server, where a uniformly trained model is used to predict faults. However, this model has inherent and insurmountable drawbacks: firstly, the geographically dispersed nature of primary substation equipment and the presence of a large amount of commercially and technically sensitive data make centralized transmission and storage highly susceptible to data leakage risks, failing to meet stringent requirements for data security and privacy protection; secondly, significant differences exist in equipment models, operating environments, and maintenance strategies among different substations. A single centralized model cannot adapt to the diverse operating conditions of each substation, resulting in insufficient accuracy in identifying complex faults, frequent false alarms and missed alarms, and difficulty in ensuring the safe and efficient operation of substations. Therefore, there is an urgent need for a technical solution that can effectively adapt to the differentiated characteristics of multiple substations while strictly protecting data privacy, thereby achieving accurate fault early warning to address the pressing needs of safe operation and maintenance of primary substation equipment. Summary of the Invention

[0003] The purpose of this invention is to address the aforementioned problems in the existing technology by providing an AI-based method, system, device, and medium for early warning of faults in substation primary equipment that can effectively adapt to the differentiated characteristics of multiple substations while strictly protecting data privacy, and achieve accurate fault warning.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] In a first aspect, the present invention provides a method for early warning of faults in primary power equipment based on artificial intelligence, the method comprising:

[0006] Edge computing nodes are deployed on multiple primary equipment in a substation, and local fault early warning models are configured on each of the multiple edge computing nodes. The local fault early warning models are used to implement graded fault early warning based on power plant operation data.

[0007] Each edge computing node independently trains a local fault early warning model based on locally collected power plant operation data. The operation data includes equipment parameters, environmental parameters, historical fault records, instantaneous disturbance data of extreme weather, equipment electromagnetic radiation intensity data, and real-time computing power data of edge nodes.

[0008] A federated learning framework is constructed, in which each edge computing node calculates its own device health correction coefficient and uploads the model parameter update information after completing local model training to the central server.

[0009] The central server aggregates and generates a global fault early warning model based on the model parameter update information received from each edge computing node, combined with an adaptive weight allocation mechanism that includes device topology correlation, computing power fluctuation factor, and device health correction coefficient.

[0010] By introducing knowledge distillation technology for cross-device interference coupling, the knowledge of the global fault warning model is transferred to the local fault warning model of each edge computing node;

[0011] Each edge computing node uses the updated local fault warning model to analyze the real-time collected power plant operation data and generate hierarchical warning information.

[0012] The calculation steps for the equipment health correction coefficient are as follows: construct an equipment health assessment model, which is used to generate the equipment health correction coefficient through multi-dimensional feature fusion. The multi-dimensional features include equipment running time, real-time temperature fluctuation, electromagnetic interference, and extreme weather.

[0013] The equipment health correction coefficient The calculation formula is:

[0014] ;

[0015] ;

[0016] in, This represents the instantaneous energy disturbance value during extreme weather events. This is the cumulative factor of equipment aging. , This represents the initial aging value of the equipment. For temperature sensitivity coefficient, For device runtime; For cross-device electromagnetic interference coupling, For equipment electromagnetic radiation power, For equipment Electromagnetic radiation power; For equipment With equipment The physical distance; , , This is the adjustment coefficient; It is a non-linear correction exponent.

[0017] The calculation steps for the equipment topology correlation degree include: firstly, constructing a topology graph based on the physical connection relationship of power plant equipment. The topology graph is stored using an adjacency matrix and features are extracted using a graph neural network. The weight of each edge in the topology graph is the fault propagation probability between the corresponding two devices. The fault propagation probability is calculated by combining historical co-occurring fault data with a time decay factor. The time decay factor decreases exponentially with the increase of the interval between the occurrence of faults in the two devices.

[0018] In the adaptive weight allocation mechanism that includes device topology correlation, computing power fluctuation factor, and device health correction coefficient, edge computing nodes In the Federal Learning Adaptive weights in round iteration The calculation formula is:

[0019] ;

[0020] ;

[0021] ;

[0022] in, For edge computing nodes The basic weights; For edge computing nodes The corresponding equipment health correction factor; These are sensitivity adjustment parameters; For edge computing nodes Mean device topology correlation within the coverage area; For edge computing nodes The computing power fluctuation factor; For the first The weight of each evaluation indicator; , For edge computing nodes Edge computing nodes In the In the first iteration Standardized scores for each evaluation indicator; For edge computing nodes The covered Topological correlation of devices; The current CPU frequency, Maximum CPU frequency; This is the difference between the CPU core temperature and the reference temperature. This is the temperature influence coefficient; To assess the number of indicators; Calculate the number of nodes at the edge.

[0023] In the knowledge distillation technique that introduces cross-device interference coupling, the distillation loss function... The calculation formula is:

[0024] ;

[0025] ;

[0026] ;

[0027] in, These are the weighting coefficients; This is the soft label loss; This represents the softening probability output by the teacher model; The softening probability output for the student model; Total number of fault categories; This is the hard label loss; One-hot encoding of the real label. The original predicted probabilities for the student model; For edge computing nodes The coupling penalty coefficient; For edge computing nodes In the Federal Learning Weights in round iteration; For edge computing nodes Cross-device electromagnetic interference coupling; The squared L2 norm of the softening probability output by the teacher model and the student model.

[0028] The real-time computing power data of the edge nodes includes: CPU utilization, memory usage, GPU core temperature, and data processing latency of the edge computing nodes.

[0029] Secondly, this invention provides an artificial intelligence-based substation primary equipment fault early warning system. The system includes a central server, a data acquisition module, a communication network, a local model training module, multiple edge computing nodes, and an early warning display platform. The multiple edge computing nodes are deployed on the substation primary equipment, and each edge computing node is configured with a local fault early warning model. The local model training module includes an equipment health calculation submodule and a model parameter update submodule. The data acquisition module collects local substation operating data, including equipment parameters, environmental parameters, historical fault records, instantaneous disturbance data from extreme weather, equipment electromagnetic radiation intensity data, and real-time computing power data of the edge nodes. The equipment health calculation submodule calculates the equipment health correction coefficient, and the model parameter update submodule updates the substation operating data collected locally at each edge computing node. A locally trained fault early warning model is used to dynamically update model parameters, and the updated model parameter information is uploaded to a central server. The central server is communicatively connected to each edge computing node. The central server includes a parameter analysis module, a global model aggregation module, and a knowledge distillation management module. The parameter analysis module is used to calculate the equipment topology correlation degree and computing power fluctuation factor. The global model aggregation module is used to receive the model parameter update information uploaded by each edge computing node, and combine it with an adaptive weight allocation mechanism that includes equipment topology correlation degree, computing power fluctuation factor, and equipment health correction coefficient to aggregate and generate a global fault early warning model. The knowledge distillation management module is used to transfer the knowledge of the global fault early warning model to the local fault early warning model of each edge computing node through knowledge distillation technology. Each edge computing node uses the updated local fault early warning model to analyze the real-time collected power plant operation data and generate hierarchical early warning information.

[0030] Thirdly, the present invention provides a substation primary equipment fault early warning device based on artificial intelligence, the substation primary equipment fault early warning device including a memory and a processor; the memory is used to store computer program code and transmit the computer program code to the processor; the processor is used to execute the substation primary equipment fault early warning method as described above according to the instructions in the computer program code.

[0031] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for early warning of faults in primary substation equipment.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] 1. The substation primary equipment fault early warning method of this invention, on the one hand, solves the data privacy and model adaptability problems of centralized training by combining distributed edge computing and federated learning, realizing localized data processing and global knowledge collaboration. Edge computing nodes are deployed in each substation, independently training local models and avoiding raw data transmission; the federated learning framework only uploads model parameter updates, combined with equipment health correction to ensure parameter quality. This edge computing and federated learning can fundamentally avoid data leakage and meet privacy protection requirements. On the other hand, adaptive weight allocation incorporates topological association, computing power factor, and equipment health correction coefficient, making the global model aggregation more in line with actual working conditions; knowledge distillation technology introduces cross-equipment interference coupling degree, balancing global generalization and local adaptability; through this multi-factor fusion adaptive weight and knowledge distillation technology, the model can adapt to the differentiated environment of different substations and absorb global experience, improving the accuracy of fault early warning; the fault propagation prediction and hierarchical early warning mechanism provides accurate decision support for operation and maintenance, ensuring the safe and efficient operation of the substation.

[0034] 2. In the substation primary equipment fault early warning method of the present invention, the calculation of health correction coefficients is based on factors such as extreme weather disturbances, aging accumulation factors, and electromagnetic interference coupling degree. This can comprehensively cover key influencing factors. Moreover, the combination of exponential and logarithmic functions characterizes the nonlinear correlation between key influencing factors, such as the logarithmic growth characteristics of electromagnetic interference. By integrating the coupling effects of multiple factors through this nonlinear function, the one-sidedness of the simple weighting method can be solved, thereby improving the calculation accuracy of the health correction coefficient.

[0035] 3. In the substation primary equipment fault early warning method described in this invention, a basic weight is used as a benchmark, and an equipment health correction coefficient is introduced to reflect the impact of equipment status on data quality; the mean of topology correlation degree reflects the system correlation of the equipment covered by the node, ensuring that the weight matches the overall structure of the power station; a computing power fluctuation factor is introduced to quantify the node processing capability through CPU frequency and temperature, avoiding interference from low-quality parameters of nodes with insufficient computing power; through this multi-factor collaborative weight calculation method, the node contribution assessment is more comprehensive, the weight of high-quality nodes is reasonably increased, and the influence of low-quality nodes is suppressed; the weight dynamically adapts to equipment status, topology correlation, and computing power changes, ensuring that the global model aggregation is more in line with the actual working conditions of the power station; the generalization ability of the global model is enhanced, and it can better adapt to the differences in equipment and environmental changes in different power stations. Attached Figure Description

[0036] Figure 1 This is a flowchart of the substation primary equipment fault early warning method described in this invention.

[0037] Figure 2 This is a schematic diagram of the substation primary equipment fault early warning system described in this invention.

[0038] Figure 3 This is a schematic diagram of the structure of the substation primary equipment fault early warning device described in this invention. Detailed Implementation

[0039] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0040] Example 1:

[0041] See Figure 1 A method for early warning of faults in primary equipment of substations based on artificial intelligence, comprising the following steps:

[0042] S1. Deploy edge computing nodes on multiple geographically dispersed primary substation equipment, and configure a local fault early warning model on each edge computing node. The local fault early warning model is used to realize hierarchical fault early warning based on substation operation data.

[0043] S2. Each edge computing node independently trains a local fault early warning model based on locally collected power plant operation data. The operation data includes equipment parameters, environmental parameters, historical fault records, instantaneous disturbance data of extreme weather, electromagnetic radiation intensity data of equipment, and real-time computing power data of edge nodes.

[0044] Specifically, the real-time computing power data of the edge nodes includes: CPU utilization, memory usage, GPU core temperature, and data processing latency of the edge computing nodes. This real-time computing power data is collected at fixed intervals by the performance monitoring module built into the edge nodes. The collected data is filtered using a sliding window to remove instantaneous fluctuations. The size of the filtering window is dynamically adjusted according to the current load of the node; the higher the load, the larger the filtering window.

[0045] S3. Construct a federated learning framework, where each edge computing node calculates its own device health correction coefficient and uploads the updated model parameters after local model training to the central server.

[0046] Specifically, traditional equipment health assessments consider only a single indicator, resulting in a one-sided assessment that fails to reflect the true state of the equipment, leading to inaccurate model parameters and impacting the reliability of early warnings. The dynamic equipment health correction method described in this invention constructs an equipment health assessment model based on equipment runtime, real-time temperature fluctuations, electromagnetic interference, and extreme weather effects. It generates equipment health correction coefficients through the fusion of multi-dimensional features. Specifically, equipment runtime reflects natural aging, real-time temperature fluctuations reflect environmental stress, electromagnetic interference is related to electromagnetic effects, and extreme weather effects encompass sudden impacts. By considering the coupled effects of these four dimensions—running time, real-time temperature fluctuations, electromagnetic interference, and extreme weather—the method covers the main influencing factors of equipment aging, avoids the limitations of a single indicator, and improves the calculation accuracy of the equipment health coefficient.

[0047] Specifically, the multi-dimensional feature fusion can employ a combination of weighted summation and nonlinear mapping. The weights for weighted summation are determined based on historical fault repair data and lifespan degradation curves provided by equipment manufacturers, ensuring that the calculated equipment health coefficient more closely matches actual and theoretical expectations. The nonlinear mapping uses a hyperbolic tangent function to normalize the fused features to the [-1, 1] interval, eliminating dimensional differences and ensuring that the obtained health correction coefficients can effectively play a role in subsequent models, providing a reliable foundation for federated learning parameter aggregation.

[0048] Specifically, the device health correction coefficient The calculation formula is:

[0049] ;

[0050] ;

[0051] ;

[0052] in, The instantaneous energy disturbance value under extreme weather conditions is collected in real time by high-frequency meteorological sensors, which collects the instantaneous energy impact of extreme weather such as thunderstorms and strong winds on the equipment, and is used to reflect the short-term impact of extreme weather on the equipment. This refers to the cumulative factor of equipment aging. This is the initial aging value of the equipment, determined by the factory parameters provided by the manufacturer. This is the temperature sensitivity coefficient, used to reflect real-time temperature fluctuations; its value is obtained by fitting long-term operating data. For device runtime, quadratic term It reflects the accelerated aging characteristics of equipment over time, and is more in line with the actual aging process; Electromagnetic interference coupling degree across devices, used to reflect the impact of electromagnetic interference. Using the inverse square relationship, the attenuation characteristics of electromagnetic interference with distance can be reflected. Summation can cover all device pairs, comprehensively quantifying the complex interactions of the electromagnetic environment within the system; For equipment electromagnetic radiation power, For equipment Electromagnetic radiation power; For equipment With equipment The physical distance; , , All of these are adjustment coefficients, which are used to balance the impact of extreme weather, equipment aging, and electromagnetic interference. They are determined through correlation analysis between historical fault data and health status. It is a nonlinear correction index, obtained by fitting historical fault data, used to adjust the nonlinear influence of aging factors, so that the calculation formula is more in line with the aging characteristics of different equipment.

[0053] Equipment health correction coefficient In the calculation formula, multi-factor coupled modeling enables the correction coefficient to truly reflect the health status of the equipment under the influence of the comprehensive environment; an exponential function is used. With Logistic function The combination of factors maps the linear combination of multiple factors to the 0-1 numerical range, thereby normalizing the health correction coefficient. The closer the value is to 1, the better the device's health status; The closer the value is to 0, the worse the health status. This nonlinear function processing can capture the complex correlation between factors and avoid the error of linear superposition. The adjustment coefficient and nonlinear correction index based on historical fault data can ensure the applicability of the formula, provide an accurate basis for the correction of model parameters in subsequent federated learning, improve the sensitivity of the fault warning model to the actual status of the equipment, and reduce false alarms or missed alarms caused by inaccurate health assessment.

[0054] S4. The central server, based on the model parameter update information received from each edge computing node, and combined with an adaptive weight allocation mechanism that includes device topology correlation, computing power fluctuation factor, and device health correction coefficient, aggregates and generates a global fault early warning model.

[0055] Specifically, traditional fault early warning models do not consider the physical connections between power plant equipment, only assessing the condition of individual devices. This ignores the transmission characteristics of faults between power plant equipment, leading to an inability to predict the scope of fault propagation, a lack of systematic early warning information, and difficulty for maintenance personnel to formulate global response strategies. This invention addresses the problem of missing fault propagation prediction caused by neglecting inter-device connections by constructing a topology graph containing fault propagation probabilities.

[0056] The steps for constructing the fault propagation topology graph include: constructing a topology graph based on the physical connection relationships of power plant equipment, such as line connections and energy transmission paths, ensuring that the correlation conforms to the actual situation. The weight of each edge in the topology graph is the fault propagation probability between the corresponding two devices, which is calculated by the frequency of the two devices failing sequentially in historical co-occurring fault data and the time decay factor. The time decay factor decreases exponentially with the increase of the interval between the failures of the two devices, that is, the longer the interval, the smaller the influence between the two devices, making the fault propagation probability more consistent with the timeliness characteristics of fault propagation. The topology graph is stored using an adjacency matrix and features are extracted using a graph neural network. The adjacency matrix storage is convenient for computer processing, and the graph neural network is used to extract topological features, which can capture the indirect correlation between devices. For example, if device A affects device B, and device B affects device C, then device A indirectly affects device C, thus enriching the feature dimensions.

[0057] This invention utilizes the correlation between devices by quantifying the probability of fault propagation through topology graphs, thereby extending early warning from a single device to the system level; it improves the timeliness of propagation probability through a time decay factor, ensuring sensitivity to recent fault correlations; and it integrates topological features extracted by graph neural networks into the topology graph, enabling the topology graph to predict fault propagation trends, providing a global perspective for operation and maintenance, facilitating the early deployment of protective measures, and reducing losses caused by fault chain reactions.

[0058] Specifically, traditional edge computing nodes, when training models, do not consider the impact of fluctuations in their own computing power, such as CPU utilization and temperature changes, on the training process. They only use fixed parameters to process data, which can easily lead to data processing delays or distortions when computing power is insufficient, affecting the training quality of local models and thus reducing the overall model aggregation effect. This invention addresses the problem of unstable data processing caused by computing power fluctuations by monitoring computing power data in real time, sensing changes in node processing capabilities, and dynamically adjusting filtering parameters accordingly, thus ensuring the quality of model training. A performance monitoring module collects CPU, memory, and GPU temperature and latency data at fixed intervals to comprehensively reflect the node's computing power status. A sliding window filter is used to smooth out instantaneous fluctuations, such as sudden CPU utilization spikes, avoiding interference from abnormal data. The window size is dynamically adjusted according to the load; a large window is used under high load to enhance the filtering effect and cope with continuous insufficient computing power, while a small window is used under low load to preserve data details and ensure sensitivity. Dynamic sliding window filtering ensures that the computing power data input to the model is stable and reliable, avoiding training interruptions or parameter deviations caused by computing power fluctuations. Ultimately, stable local data processing provides high-quality input for model training, indirectly improving the accuracy of the local model and the subsequent global model aggregation effect, ensuring that the entire early warning system can still operate stably when the computing power of edge nodes changes dynamically.

[0059] Specifically, traditional federated learning often relies on sample size or evaluation metrics for weight allocation, failing to consider equipment topology correlation and edge node computing power fluctuations. This results in weights that don't reflect the actual contribution of nodes, underestimating the weights of high-quality nodes, and allowing low-quality nodes to interfere with the global model. Consequently, the aggregated global model lacks generalization ability and struggles to adapt to the complex operating conditions of substation primary equipment. This invention addresses the limitations of traditional weight allocation, which relies solely on sample size or evaluation metrics, by dynamically calculating the aggregation weights of each edge node in federated learning. By integrating the basic sample size, equipment health correction coefficients, equipment topology correlation, and computing power status, it ensures that high-quality nodes play a greater role in the global model aggregation.

[0060] In the adaptive weight allocation mechanism that includes device topology correlation, computing power fluctuation factor, and device health correction coefficient, edge computing nodes In the Federal Learning Adaptive weights in round iteration The calculation formula is:

[0061] ;

[0062] ;

[0063] ;

[0064] in, For edge computing nodes The basic weights are determined by the edge nodes. The number of local training samples, sample balance, and uniformity of sample distribution across different fault types are all factors that determine the effectiveness of the test. A larger sample size and a more balanced distribution result in better performance. The larger the value, the more it reflects the fundamental contribution to the scale and quality of the data; For edge computing nodes The corresponding device health correction coefficient is higher for nodes with higher health. This value helps avoid low-quality data from poorly functioning equipment affecting calculation accuracy. These are sensitivity adjustment parameters; For edge computing nodes The average value of the topological correlation of devices within the coverage area is averaged to reflect the overall correlation strength between devices within the coverage area. The more closely related the nodes are in the topology, the more significant the weight is, which is consistent with the system characteristics of fault propagation. For edge computing nodes The computing power fluctuation factor; For the first The weights of the evaluation indicators are determined by combining the analytic hierarchy process with backpropagation of fault early warning errors. , For edge computing nodes Edge computing nodes In the In the first iteration Standardized scores for each evaluation indicator; For edge computing nodes The covered Topological correlation of devices; The current CPU frequency, Maximum CPU frequency; This is the difference between the CPU core temperature and the reference temperature. This is the temperature influence coefficient; the higher the temperature, the greater the influence. The smaller, the more Reduce the excessive participation of nodes whose computing power is reduced due to high temperatures in aggregation; To assess the number of indicators; Calculate the number of nodes at the edge.

[0065] The above adaptive weights In the calculation formula, the basic weights are used. Based on this, a device health correction factor is introduced. Reflects the impact of device status on data quality; through the mean of topological correlation degree. This reflects the system correlation of node coverage equipment, ensuring that the weights match the overall structure of the power plant; and introduces a computing power fluctuation factor. By quantifying node processing capabilities through CPU frequency and temperature, interference from low-quality parameters of nodes with insufficient computing power can be avoided. For edge computing nodes The overall evaluation score, The sum of the comprehensive evaluation scores for all edge computing nodes is calculated by weighting the evaluation indicators in the numerator and denominator to ensure the relative reasonableness of the weights. The index term controls the sensitivity to the overall score, ensuring that the weight of high-quality nodes is appropriately amplified. In summary, the multi-factor collaborative weight calculation method adopted in this invention makes the evaluation of node contributions more comprehensive, reasonably increases the weight of high-quality nodes, and suppresses the impact of low-quality nodes; the weights dynamically adapt to equipment status, topology associations, and computing power changes, ensuring that the global model aggregation better fits the actual operating conditions of the power plant; the generalization ability of the global model is enhanced, enabling it to better adapt to equipment differences and environmental changes in different power plants.

[0066] This step also incorporates adaptive weights, including device topology correlation and computing power fluctuation factors, based on the received model parameter update information for each edge computing node. This aggregates and generates a global fault early warning model.

[0067] S5. By introducing knowledge distillation technology for cross-device interference coupling, the knowledge of the global fault warning model is transferred to the local fault warning model of each edge computing node.

[0068] Specifically, traditional knowledge distillation techniques only transfer global knowledge through soft-label and hard-label losses, without considering the impact of cross-device electromagnetic interference on the local model. This results in insufficient generalization ability of the local model in interference environments, making it difficult to balance global knowledge with local anti-interference requirements, and causing a significant decrease in warning accuracy in complex electromagnetic environments. This invention introduces an electromagnetic interference coupling penalty term on top of the traditional soft-label and hard-label losses to solve the problem of insufficient generalization ability of the local model in complex electromagnetic environments, achieving coordinated transfer of global knowledge and local anti-interference capabilities.

[0069] Specifically, in the knowledge distillation technique that introduces cross-device interference coupling, the distillation loss function used is... The calculation formula is:

[0070] ;

[0071] ;

[0072] ;

[0073] in, These are weighting coefficients that can be dynamically adjusted with each training round, for example, in the early stages of training. Higher values ​​are prioritized for learning global knowledge, especially in the later stages of training. The value is reduced to enhance local adaptation, thereby balancing the transfer ratio of global and local knowledge; For soft label loss, KL divergence is used to measure the difference in the soft label probability distribution between the student model and the teacher model; The teacher model represents the fault category. Output softening probability; For the student model, for fault categories Output softening probability; Total number of fault categories; For hard label loss, cross-entropy is used to measure the difference between the student model's predicted values ​​and the true labels; One-hot encoding of the real label. The original predicted probabilities for the student model; For edge computing nodes The coupling penalty coefficient; Its value is negatively correlated with the equipment health correction coefficient; equipment with a lower health correction coefficient corresponds to a higher value. Value, specifically enhance the anti-interference training intensity of such weak equipment; For edge computing nodes In the Federal Learning Weights in round iteration; For edge computing nodes Cross-device electromagnetic interference coupling; For the teacher model and student model regarding fault categories The squared L2 norm of the output softening probability.

[0074] In the distillation loss function In the middle, through and To ensure both global knowledge transfer and local data fitting, a new interference coupling penalty term is added. This allows the local model to adapt to the electromagnetic interference environment while learning global knowledge, improving the stability of early warning in complex scenarios; dynamically adjusted and Balancing global generalization with local anti-interference requirements, avoiding model overfitting to a single scenario; the distilled local model can both absorb global experience and resist electromagnetic interference between devices, maintaining stable early warning performance in the differentiated environment of multiple power stations.

[0075] S6. Each edge computing node uses the updated local fault early warning model to analyze the real-time collected power plant operation data and generate graded fault early warning information.

[0076] Example 2:

[0077] See Figure 2An AI-based substation primary equipment fault early warning system is disclosed, applied to the AI-based substation primary equipment fault early warning method described in Example 1. The system includes a central server, a data acquisition module, a communication network, a local model training module, multiple edge computing nodes, and an early warning display platform. The multiple edge computing nodes are deployed on the substation primary equipment, and each edge computing node is configured with a local fault early warning model. The local model training module includes an equipment health calculation submodule and a model parameter update submodule. The data acquisition module, deployed on each substation primary equipment, includes equipment sensor groups, environmental monitoring sensor groups, electromagnetic radiation sensor groups, and edge node performance sensor groups, used to collect real-time substation operating data. The substation operating data includes equipment parameters, environmental parameters, historical fault records, instantaneous disturbance data from extreme weather, equipment electromagnetic radiation intensity data, and real-time computing power data of the edge nodes. The equipment health calculation submodule is used to calculate the equipment health correction coefficient. The model parameter update submodule is used to independently train the local fault early warning model based on the substation operating data and equipment health correction coefficient collected locally at each edge computing node, so as to dynamically update the model. The system generates new model parameters and uploads the updated parameters to a central server. The central server is connected to each edge computing node and includes a parameter analysis module, a global model aggregation module, and a knowledge distillation management module. The parameter analysis module calculates the device topology correlation and computing power fluctuation factor. The global model aggregation module receives the updated model parameters uploaded by each edge computing node and, using an adaptive weight allocation mechanism that includes device topology correlation and computing power fluctuation factor, aggregates and generates a global fault early warning model. The knowledge distillation management module uses knowledge distillation technology to transfer the knowledge of the global fault early warning model to the local fault early warning models of each edge computing node. Each edge computing node uses the updated local fault early warning model to analyze the real-time collected power plant operation data and generate tiered early warning information. The communication network enables secure data transmission between the edge computing nodes and the central server. The communication network uses a combination of encryption protocols and dynamic port switching to ensure data transmission security. The encryption protocol uses national cryptographic algorithms, and the dynamic port switching cycle is adjusted according to network load fluctuations. The early warning display platform displays the tiered early warning information.

[0078] The aforementioned system, through a modular collaborative architecture, addresses the shortcomings of traditional centralized systems, achieving a fusion of the advantages of distributed and centralized approaches. Edge computing nodes deploy local fault early warning models, enabling localized data processing and real-time early warning; the central server is responsible for global model aggregation and knowledge migration, coordinating multi-node collaboration; the communication network employs national cryptographic algorithms and dynamic port switching, providing dual protection for parameter transmission security; the data acquisition module covers multi-dimensional sensors, providing comprehensive input for the model; and the early warning display platform integrates various types of information, offering intuitive references for operation and maintenance. Through the collaboration of distributed edge nodes and a centralized central server, data privacy is protected while global knowledge sharing is achieved; the modular design ensures clear functionalities, efficient interaction, and strong system scalability; secure communication mechanisms eliminate the risk of parameter transmission leakage; and comprehensive data acquisition and an intuitive display platform improve the timeliness of fault early warnings and the accuracy of operation and maintenance decisions, ensuring the stable operation of primary substation equipment.

[0079] Example 3:

[0080] See Figure 3 A fault early warning device for primary substation equipment based on artificial intelligence includes a memory and a processor; the memory is used to store computer program code and transmit the computer program code to the processor; the processor is used to execute the fault early warning method for primary substation equipment described in Embodiment 1 according to the instructions in the computer program code.

[0081] Example 4:

[0082] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the substation primary equipment fault early warning method described in Embodiment 1.

[0083] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program goods. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0085] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0086] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for early warning of faults in primary equipment of substations based on artificial intelligence, characterized in that: The substation primary equipment fault early warning method includes: Edge computing nodes are deployed on multiple primary equipment in a substation, and local fault early warning models are configured on each of the multiple edge computing nodes. The local fault early warning models are used to implement graded fault early warning based on power plant operation data. Each edge computing node independently trains a local fault early warning model based on locally collected power plant operation data, which includes equipment parameters, environmental parameters, historical fault records, instantaneous disturbance data of extreme weather, equipment electromagnetic radiation intensity data, and real-time computing power data of edge nodes. A federated learning framework is constructed, in which each edge computing node calculates its own device health correction coefficient and uploads the model parameter update information after completing local model training to the central server. The central server aggregates and generates a global fault early warning model based on the model parameter update information received from each edge computing node, combined with an adaptive weight allocation mechanism that includes device topology correlation, computing power fluctuation factor, and device health correction coefficient. By introducing knowledge distillation technology for cross-device interference coupling, the knowledge of the global fault warning model is transferred to the local fault warning model of each edge computing node; Each edge computing node uses the updated local fault warning model to analyze the real-time collected power plant operation data and generate hierarchical warning information.

2. The method for early warning of substation primary equipment faults based on artificial intelligence according to claim 1, characterized in that: The calculation steps for the equipment health correction coefficient are as follows: construct an equipment health assessment model, which is used to generate the equipment health correction coefficient through multi-dimensional feature fusion. The multi-dimensional features include equipment running time, real-time temperature fluctuation, electromagnetic interference, and extreme weather.

3. The method for early warning of faults in primary substation equipment based on artificial intelligence according to claim 2, characterized in that: The equipment health correction coefficient The calculation formula is: ； ； in, This represents the instantaneous energy disturbance value during extreme weather events. This is the cumulative factor of equipment aging. , This represents the initial aging value of the equipment. For temperature sensitivity coefficient, For device runtime; For cross-device electromagnetic interference coupling, For equipment electromagnetic radiation power, For equipment Electromagnetic radiation power; For equipment With equipment The physical distance; , , This is the adjustment coefficient; It is a non-linear correction exponent.

4. The method for early warning of substation primary equipment faults based on artificial intelligence according to any one of claims 1-3, characterized in that: The calculation steps for the equipment topology correlation degree include: firstly, constructing a topology graph based on the physical connection relationship of power plant equipment. The topology graph is stored using an adjacency matrix and features are extracted using a graph neural network. The weight of each edge in the topology graph is the fault propagation probability between the corresponding two devices. The fault propagation probability is calculated by combining historical co-occurring fault data with a time decay factor. The time decay factor decreases exponentially with the increase of the interval between the occurrence of faults in the two devices.

5. The method for early warning of faults in primary substation equipment based on artificial intelligence according to any one of claims 1-3, characterized in that: In the adaptive weight allocation mechanism that includes device topology correlation, computing power fluctuation factor, and device health correction coefficient, edge computing nodes In the Federal Learning Adaptive weights in round iteration The calculation formula is: ； ； ； in, For edge computing nodes The basic weights; For edge computing nodes The corresponding equipment health correction factor; These are sensitivity adjustment parameters; For edge computing nodes Mean device topology correlation within the coverage area; For edge computing nodes The computing power fluctuation factor; For the first The weight of each evaluation indicator; , For edge computing nodes Edge computing nodes In the In the first iteration Standardized scores for each evaluation indicator; For edge computing nodes The covered Topological correlation of devices; Current CPU frequency; Maximum CPU frequency; This is the difference between the CPU core temperature and the reference temperature. This is the temperature influence coefficient; To assess the number of indicators; Calculate the number of nodes at the edge.

6. The method for early warning of faults in primary substation equipment based on artificial intelligence according to any one of claims 1-3, characterized in that: In the knowledge distillation technique that introduces cross-device interference coupling, the distillation loss function... The calculation formula is: ； ； ； in, These are the weighting coefficients; This is the soft label loss; This represents the softening probability output by the teacher model; The softening probability output for the student model; Total number of fault categories; This is the hard label loss; One-hot encoding of the real label. The original predicted probabilities for the student model; For edge computing nodes The coupling penalty coefficient; For edge computing nodes In the Federal Learning Weights in round iteration; For edge computing nodes Cross-device electromagnetic interference coupling; The squared L2 norm of the softening probability output by the teacher model and the student model.

7. The method for early warning of substation primary equipment faults based on artificial intelligence according to any one of claims 1-3, characterized in that: The real-time computing power data of the edge nodes includes: CPU utilization, memory usage, GPU core temperature, and data processing latency of the edge computing nodes.

8. A fault early warning system for primary equipment in substations based on artificial intelligence, characterized in that: The substation primary equipment fault early warning system includes a central server, a data acquisition module, a communication network, a local model training module, multiple edge computing nodes, and an early warning display platform. The multiple edge computing nodes are deployed on the substation primary equipment, and each edge computing node is configured with a local fault early warning model. The local model training module includes an equipment health calculation submodule and a model parameter update submodule. The data acquisition module collects local substation operating data, including equipment parameters, environmental parameters, historical fault records, instantaneous disturbance data from extreme weather, equipment electromagnetic radiation intensity data, and real-time computing power data from the edge nodes. The equipment health calculation submodule calculates the equipment health correction coefficient, and the model parameter update submodule updates the model parameters based on the data collected locally at each edge computing node. The system independently trains local fault early warning models using the collected power plant operation data to dynamically update model parameters and uploads the updated model parameter information to a central server. The central server is communicatively connected to each edge computing node and includes a parameter analysis module, a global model aggregation module, and a knowledge distillation management module. The parameter analysis module is used to calculate the equipment topology correlation degree and computing power fluctuation factor. The global model aggregation module is used to receive the model parameter update information uploaded by each edge computing node and, combined with an adaptive weight allocation mechanism that includes equipment topology correlation degree, computing power fluctuation factor, and equipment health correction coefficient, aggregates and generates a global fault early warning model. The knowledge distillation management module is used to transfer the knowledge of the global fault early warning model to the local fault early warning models of each edge computing node through knowledge distillation technology. Each edge computing node uses the updated local fault warning model to analyze the real-time collected power plant operation data and generate hierarchical warning information.

9. A fault early warning device for primary substation equipment based on artificial intelligence, characterized in that: The substation primary equipment fault early warning device includes a memory and a processor; the memory is used to store computer program code and transmit the computer program code to the processor; the processor is used to execute the substation primary equipment fault early warning method as described in claim 1 according to the instructions in the computer program code.

10. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the substation primary equipment fault early warning method as described in claim 1.

Images