Circuit breaker fault monitoring system and method based on artificial intelligence

Abstract

The invention discloses a circuit breaker fault monitoring system and method based on artificial intelligence, and relates to the technical field of fault monitoring, and the method comprises the steps: calculating and generating different key physical models through the physical data of a circuit breaker, outputting the waveform data of the different key physical models, and forming a simulation data set; extracting abstract features from the waveform data by using a signal processing algorithm, and constructing a mixed feature space by using the physical features and the abstract features; training a circuit breaker fault prediction model by using a neural network algorithm, and forming a total loss function by combining prediction error loss with physical residual error loss; extracting a training sample in the simulation data set, and performing meta-learning on the fault prediction model; real-time physical data of the circuit breaker is extracted and mapped to the mixed feature space, the similar distance between the real-time physical data and each prototype is calculated, and the real-time fault type is judged; constructing a knowledge graph according to expert experience; and searching the dimension action of the real-time fault type and the fault reason in the knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of fault monitoring technology, specifically to a circuit breaker fault monitoring system and method based on artificial intelligence. Background Technology

[0002] Power systems are crucial in modern society, supporting all aspects of industrial production, commercial operations, and residential life. High-voltage circuit breakers, as key equipment in power systems, bear the important responsibility of controlling and protecting circuits. They can connect and disconnect load currents under normal conditions and quickly interrupt short-circuit currents in the event of a fault, protecting other equipment in the power system from damage and preventing the fault from escalating. With the continuous expansion of power system scale and the continuous increase in voltage levels, the number of high-voltage circuit breakers is also increasing. Traditional fault diagnosis methods mainly rely on manual inspection and experience-based judgment, which suffers from low efficiency, poor accuracy, and insufficient real-time performance, making it difficult to meet the fault diagnosis needs of large-scale, high-voltage power systems. Intelligent fault diagnosis technology integrates knowledge and methods from multiple disciplines such as sensor technology, signal processing technology, data analysis technology, and artificial intelligence technology. It can monitor and analyze the operating status of high-voltage circuit breakers in real time, diagnose fault types and locations accurately and promptly, and provide corresponding fault handling suggestions. This enables condition-based maintenance of high-voltage circuit breakers, avoiding the blindness and over-maintenance of traditional periodic maintenance, effectively improving equipment operational reliability, reducing operation and maintenance costs, and enhancing the overall operating efficiency of the power system.

[0003] However, in traditional circuit breaker fault monitoring, when a brand-new circuit breaker or a newly emerging fault mode lacks historical data accumulation, traditional data-driven models cannot be trained or achieve extremely poor diagnostic results. A model trained on one operating condition or type of circuit breaker may show a sharp decline in performance on a slightly different operating condition or type, lacking universality. Furthermore, when purely data-driven deep learning models provide diagnostic results, they lack physical mechanism support and logical explanation, making them difficult for field engineers and domain experts to accept and trust. For special faults with extremely low probability of occurrence, due to the scarcity of samples, traditional models often treat them as noise and ignore them, leading to missed detections. Summary of the Invention

[0004] The purpose of this invention is to provide a circuit breaker fault monitoring system and method based on artificial intelligence to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A circuit breaker fault monitoring method based on artificial intelligence, the method comprising the following steps: S100. Calculate and generate different key physical models using the physical data of the circuit breaker, collect physical data of the health and fault states of the circuit breaker in history, input them into the key physical models for simulation, output waveform data of different key physical models, and form a simulation dataset. Furthermore, the specific steps for forming the simulation dataset are as follows: S101. The key physical model includes an electromagnetic operating mechanism model, a mechanical motion model, and a contact electrical contact model. The specific construction method of the electromagnetic operating mechanism model is as follows: collect the coil current, coil voltage, electromagnetic force, and moving iron core displacement during circuit breaker operation; and construct the influence relationship between coil current, electromagnetic force, and moving iron core displacement on coil voltage based on the physical principles of circuit breaker operation. The formula is: ; In the formula, U represents the coil voltage, L represents the coil inductance, i represents the coil current, R represents the coil DC resistance, and K represents the coil resistance. e Let x represent the back electromotive force constant, x represent the displacement of the moving iron core, and t represent time. Based on actual conditions, professionals derive the relationship between the coil current and the displacement of the moving iron core on the electromagnetic force as F. magnetic =f(i,x), where F magnetic Let f represent electromagnetic force and f represent the functional relationship; a model of an electromagnetic operating mechanism is constructed using these two influence relationships. S102. The specific method for constructing the mechanical motion model is as follows: Analyze the forces acting on the circuit breaker during operation, collect the spring reaction force, friction force, damping force, and total mass when driving the iron core, and calculate the resultant force of the acceleration generated by the total mass of the circuit breaker. The formula is: ; In the formula, m represents the total mass, and F spring F represents the spring reaction force. friction F represents frictional force. damping Indicates damping force. The second derivative of the displacement of the moving iron core with respect to time is used; the formula for calculating the resultant force is used as a model of mechanical motion. S103. Professionals actively construct an electrical contact model of the contact based on the relationship between contact resistance and contact pressure, material and degree of burning. The key physical models encompass the electromagnetic operating mechanism model, the mechanical motion model, and the contact electrical contact model, each corresponding to a core operating component of the circuit breaker. The electromagnetic operating mechanism model, built upon key physical quantities such as coil voltage, current, electromagnetic force, and moving core displacement, accurately reflects the operating state of the electromagnetic drive section. The mechanical motion model accurately describes the mechanical transmission process by analyzing the effects of spring reaction force, friction, and other forces. The contact electrical contact model focuses on the relationship between contact resistance, contact pressure, and materials, capturing the key influencing factors of conductive contact. These three models comprehensively and accurately characterize the circuit breaker's operating mechanism, laying a solid foundation for subsequent fault monitoring.

[0006] The physical data of the historical circuit breaker health status and different fault states are collected and input into three models to obtain waveform data to form a simulation dataset.

[0007] Historical physical data on the health status and various fault states of circuit breakers are collected and input into a key physical model to generate a simulation dataset. This dataset includes baseline data for normal circuit breaker operation as well as data under multiple fault states, comprehensively simulating various operating scenarios of circuit breakers. Subsequent model training and feature extraction based on this dataset allow the model to fully learn the feature differences under different states, greatly improving the model's fault identification ability and reducing the problem of poor model generalization caused by insufficient actual data or a single scenario.

[0008] S200: Extract physical features related to the key physical model during circuit breaker operation, extract abstract features from waveform data using signal processing algorithms, and construct a hybrid feature space using physical features and abstract features. Furthermore, the specific steps for constructing a hybrid feature space using physical and abstract features are as follows: S201. Collect all working data involved in the operation of the circuit breaker. When the circuit breaker fails, calculate the correlation coefficient between each type of working data and each type of physical data in the key physical model. Select the working data with a correlation coefficient of non-zero as the physical features related to the key physical model of the circuit breaker. S202. Use wavelet transform algorithm to extract frequency domain features from waveform data in simulation dataset, use frequency domain features as abstract features of circuit breaker, and use abstract features and physical features together to form a hybrid feature space.

[0009] Wavelet transform algorithms are used to extract frequency domain features from waveform data as abstract features. Wavelet transform offers multi-resolution analysis capabilities, effectively capturing the variation characteristics of waveform data across different frequency ranges. These frequency domain features often contain potential information about circuit breaker faults. For example, certain faults may cause signal enhancement or attenuation at specific frequency components. By using abstract features, this fault information, which is not easily observed directly, can be extracted, complementing the physical features.

[0010] Physical characteristics can intuitively reflect the macroscopic changes in the operating state of a circuit breaker, while abstract characteristics can uncover microscopic fault information hidden in the data. The hybrid feature space constructed by combining the two takes into account both macroscopic and microscopic, direct and indirect feature information, and can more comprehensively describe the state of the circuit breaker. Fault prediction and type determination based on this feature space can significantly improve the accuracy and comprehensiveness of fault identification, and reduce misjudgments or omissions caused by single features.

[0011] S300. A circuit breaker fault prediction model is trained using a neural network algorithm. An additional physical residual loss is designed into the prediction error loss constructed during the training of the neural network algorithm. The prediction error loss and the physical residual loss are combined to form the total loss function. Training samples are extracted from the simulation dataset to perform meta-learning on the fault prediction model. Furthermore, the specific steps for constructing the total loss function by combining the prediction error loss with the physical residual loss are as follows: S301. A circuit breaker fault prediction model is trained using a neural network algorithm. The prediction error loss constructed during the training of the neural network algorithm is L. date The prediction error loss is used to measure the error between the predicted result output by the fault prediction model and the true value; an additional physical residual loss is designed into the prediction error loss constructed during the training of the neural network algorithm, with the following formula: ; In the formula, L physics Represents physical residual loss, Physics law Representing different key physical models, f θ (s) represents the different physical prediction values ​​of the fault prediction model when predicting circuit breaker faults, y physics This indicates that the predicted physical value corresponds to the observed actual physical value. S302. Combine the prediction error loss and the physical residual loss to form the total loss function, as shown in the formula: ; In the formula, L total Let λ represent the total loss and λ represent the hyperparameter. By combining prediction error loss and physical residual loss, and adjusting their weights through the hyperparameter λ, the model's prediction accuracy is guaranteed while adhering to physical laws. This avoids situations where the model violates physical principles in pursuit of high prediction accuracy, making model training more scientific and reasonable.

[0012] S303. Extract different states of the circuit breaker from the simulation dataset, providing k samples for each state to form a support set. Then extract n samples as a query set. Combine the support set and the query set to form a task. Use model-independent meta-learning to learn the support set and the query set. Set up an inner loop and an outer loop. In the inner loop, extract a task and use the initial parameters of the fault prediction model, which represent the initial values ​​of different model parameters in the fault prediction model. Perform forward propagation on the support set, calculate the total loss, calculate the gradient of the total loss with respect to the initial parameters, subtract the gradient from the initial parameters, and repeat the inner loop to obtain the adaptation parameters of the fault prediction model after adapting to the task. In the inner loop, the model is trained on the support set using initial parameters. By calculating the total loss and gradients and updating the parameters, it quickly adapts to the current task. This process enables the model to rapidly adjust its parameters when faced with new tasks (different combinations of states), improving its adaptability to specific tasks.

[0013] In the outer loop, the adapted fault prediction model is used to propagate forward on the query set for each task, and the total loss is calculated. The total loss of the query set for all tasks is summed to form the meta-loss. The gradient of the meta-loss with respect to the initial parameters is calculated, and the optimized parameters are obtained by subtracting the gradient from the initial parameters. In the meta-learning, the outer loop optimizes the initial parameters of the fault prediction model, and the inner loop uses the optimized parameters as the initial parameters to perform an adaptation task to obtain the adapted parameters.

[0014] The outer loop summarizes the total loss of all tasks' query sets to form the meta-loss. By optimizing the initial parameters, the model's initial parameters are made more generalizable, enabling it to quickly adapt to more diverse tasks. The fault prediction model trained through meta-learning can still maintain high prediction accuracy when facing new fault types or new working scenarios that have not been seen in reality, significantly improving the model's generalization ability and reducing its limitations in practical applications.

[0015] S400. For each fault type, feature vectors are extracted in the hybrid feature space and the mean of the feature vectors is calculated to form a "prototype" of the fault type. When the fault prediction model detects a circuit breaker fault, real-time physical data of the circuit breaker is extracted and mapped to the hybrid feature space to calculate the similarity distance with each "prototype" to determine the real-time fault type. Furthermore, the specific steps for determining the type of real-time fault are as follows: S401. In the hybrid feature space, transform the features of each circuit breaker fault type into feature vectors. Calculate the average value of each feature vector for each fault type, and use the average value of each feature vector as the "prototype p" of the corresponding fault type. c ”; S402. When a circuit breaker fault is detected, the circuit breaker data output from the fault prediction model at the time of the fault is mapped into the hybrid feature space. The feature vector of the real-time fault is extracted, and the similarity distance between the real-time fault and each "prototype" of the circuit breaker is calculated. The formula is as follows: ; In the formula, Sim represents the similarity distance between the real-time fault and each "prototype" of the circuit breaker, p(s) test ) represents the feature vector of a real-time fault; the fault type corresponding to the "prototype" with the minimum similarity distance is selected as the real-time fault type.

[0016] When a circuit breaker fault is detected, real-time fault data is mapped to a hybrid feature space to extract feature vectors. The fault type is determined by calculating the similarity distance with each fault "prototype". This method, based on feature vector space distance, intuitively reflects the similarity between the real-time fault and known fault types; the smaller the distance, the higher the similarity. Selecting the fault type corresponding to the minimum similarity distance as the real-time fault type allows for quick and accurate fault type determination, saving valuable time for subsequent fault handling and reducing delays or errors caused by incorrect fault type identification.

[0017] S500: Construct a knowledge graph based on expert experience; search for dimensional actions and causes of real-time fault types in the knowledge graph.

[0018] Furthermore, the specific steps for finding the dimensional actions and causes of real-time fault types in the knowledge graph are as follows: S501. Experts actively input maintenance actions, fault causes, and correlations for different fault types, using fault type, fault cause, fault characteristics, and maintenance actions as nodes and correlations as edges, and construct a knowledge graph based on expert experience. S502. After determining the fault type of a real-time fault, locate the corresponding fault type in the knowledge graph, extract the nodes related to the real-time fault type based on the edge extraction of the knowledge graph, extract the maintenance actions and fault causes of the real-time fault pair, and push them to the maintenance personnel.

[0019] Once the real-time fault type is determined, the corresponding fault type node can be quickly located in the knowledge graph. Related fault causes and maintenance action nodes can then be extracted through the association edges. This query method eliminates the need for manual searching through large amounts of documents or data, enabling the acquisition of accurate and comprehensive fault-related information in a short time. This provides timely decision support for maintenance personnel, helping them quickly develop fault handling plans, improve fault handling efficiency, and reduce the impact of faults on the normal operation of circuit breakers.

[0020] An artificial intelligence-based circuit breaker fault monitoring system includes a data acquisition module, a feature extraction module, a model training module, a fault type lookup module, and a fault reasoning module. The data acquisition module is used to calculate and generate different key physical models using the physical data of the circuit breaker, and to collect physical data on the health status and fault status of the circuit breaker in history. The feature extraction module is used to extract physical features related to the key physical model during circuit breaker operation, extract abstract features from waveform data using signal processing algorithms, and construct a hybrid feature space using physical features and abstract features. The model training module is used to train a circuit breaker fault prediction model using a neural network algorithm, extract training samples from the simulation dataset, and perform meta-learning on the fault prediction model. The fault type lookup module is used to extract real-time physical data of the circuit breaker and map it into a hybrid feature space when the fault prediction model detects a circuit breaker fault, calculate the similarity distance with each "prototype", and determine the real-time fault type. The fault reasoning module is used to construct a knowledge graph based on expert experience; and to search for the dimensional actions and causes of real-time fault types in the knowledge graph.

[0021] The data acquisition module includes key physical model units and simulation units; The key physical model unit is used to calculate and generate different key physical models using the physical data of the circuit breaker. The key physical models include electromagnetic operating mechanism model, mechanical motion model and contact electrical contact model. The simulation unit is used to collect physical data of the historical circuit breaker health status and different fault states, and input them into three models to obtain waveform data to form a simulation dataset.

[0022] The feature extraction module includes physical feature units and abstract feature units; The physical feature unit is used to calculate the correlation coefficient between each type of working data and each type of physical data in the key physical model when the circuit breaker fails, and selects the working data with a correlation coefficient of non-zero as the physical feature related to the key physical model of the circuit breaker. The abstract feature unit is used to extract frequency domain features from waveform data in the simulation dataset using wavelet transform algorithm, and the frequency domain features are used as abstract features of the circuit breaker.

[0023] The model training module includes a total loss function construction unit and a meta-learning unit; The total loss function construction unit is used to combine the prediction error loss and the physical residual loss to form the total loss function; The meta-learning unit is used to extract training samples from the simulation dataset and perform meta-learning on the fault prediction model.

[0024] Compared with the prior art, the beneficial effects of the present invention are: 1. From the construction of key physical models and hybrid feature spaces, to the design of the total loss function and meta-learning training, and then to the establishment of fault "prototypes" and similarity distance calculation, each step revolves around improving the accuracy of fault monitoring. Through multi-dimensional and multi-level optimization, the model can accurately identify the fault state of the circuit breaker and accurately determine the fault type, significantly reducing the fault misjudgment rate and missed judgment rate, and providing strong support for the reliable operation of the circuit breaker.

[0025] 2. On the one hand, meta-learning training enhances the model's generalization ability, enabling it to adapt to different working scenarios and new fault types. On the other hand, the knowledge graph, built upon expert experience, provides fault causes and maintenance actions with strong practical guidance, allowing maintenance personnel to quickly and effectively handle faults based on the queried information. Furthermore, the simulation dataset used in the solution can be continuously updated and expanded according to actual conditions, further enhancing the adaptability and flexibility of the solution in practical applications, and meeting the fault monitoring needs of circuit breakers of different models and operating environments. Attached Figure Description

[0026] Figure 1 This is a module distribution diagram of an artificial intelligence-based circuit breaker fault monitoring system according to the present invention. Figure 2 This is a schematic diagram illustrating the steps of an artificial intelligence-based circuit breaker fault monitoring method according to the present invention. Detailed Implementation

[0027] 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.

[0028] Example: Figures 1-2 As shown, the present invention provides a technical solution. A circuit breaker fault monitoring method based on artificial intelligence, the method comprising the following steps: S100. Calculate and generate different key physical models using the physical data of the circuit breaker, collect physical data of the health and fault states of the circuit breaker in history, input them into the key physical models for simulation, output waveform data of different key physical models, and form a simulation dataset. The specific steps to create the simulation dataset are as follows: S101. The key physical model includes an electromagnetic operating mechanism model, a mechanical motion model, and a contact electrical contact model. The specific construction method of the electromagnetic operating mechanism model is as follows: collect the coil current, coil voltage, electromagnetic force, and moving iron core displacement during circuit breaker operation; and construct the influence relationship between coil current, electromagnetic force, and moving iron core displacement on coil voltage based on the physical principles of circuit breaker operation. The formula is: ; In the formula, U represents the coil voltage, L represents the coil inductance, i represents the coil current, R represents the coil DC resistance, and K represents the coil resistance. e Let x represent the back electromotive force constant, x represent the displacement of the moving iron core, and t represent time. Based on actual conditions, professionals derive the relationship between the coil current and the displacement of the moving iron core on the electromagnetic force as F. magnetic =f(i,x), where F magnetic Let f represent electromagnetic force and f represent the functional relationship; a model of an electromagnetic operating mechanism is constructed using these two influence relationships. S102. The specific method for constructing the mechanical motion model is as follows: Analyze the forces acting on the circuit breaker during operation, collect the spring reaction force, friction force, damping force, and total mass when driving the iron core, and calculate the resultant force of the acceleration generated by the total mass of the circuit breaker. The formula is: ; In the formula, m represents the total mass, and F spring F represents the spring reaction force. friction F represents frictional force. damping Indicates damping force. The second derivative of the displacement of the moving iron core with respect to time is used; the formula for calculating the resultant force is used as a model of mechanical motion. S103. Professionals actively construct an electrical contact model of the contact based on the relationship between contact resistance and contact pressure, material and degree of burning. The key physical models encompass the electromagnetic operating mechanism model, the mechanical motion model, and the contact electrical contact model, each corresponding to a core operating component of the circuit breaker. The electromagnetic operating mechanism model, built upon key physical quantities such as coil voltage, current, electromagnetic force, and moving core displacement, accurately reflects the operating state of the electromagnetic drive section. The mechanical motion model accurately describes the mechanical transmission process by analyzing the effects of spring reaction force, friction, and other forces. The contact electrical contact model focuses on the relationship between contact resistance, contact pressure, and materials, capturing the key influencing factors of conductive contact. These three models comprehensively and accurately characterize the circuit breaker's operating mechanism, laying a solid foundation for subsequent fault monitoring.

[0029] The physical data of the historical circuit breaker health status and different fault states are collected and input into three models to obtain waveform data to form a simulation dataset.

[0030] Historical physical data on the health status and various fault states of circuit breakers are collected and input into a key physical model to generate a simulation dataset. This dataset includes baseline data for normal circuit breaker operation as well as data under multiple fault states, comprehensively simulating various operating scenarios of circuit breakers. Subsequent model training and feature extraction based on this dataset allow the model to fully learn the feature differences under different states, greatly improving the model's fault identification ability and reducing the problem of poor model generalization caused by insufficient actual data or a single scenario.

[0031] S200: Extract physical features related to the key physical model during circuit breaker operation, extract abstract features from waveform data using signal processing algorithms, and construct a hybrid feature space using physical features and abstract features. The specific steps for constructing a hybrid feature space using physical and abstract features are as follows: S201. Collect all working data involved in the operation of the circuit breaker. When the circuit breaker fails, calculate the correlation coefficient between each type of working data and each type of physical data in the key physical model. Select the working data with a correlation coefficient of non-zero as the physical features related to the key physical model of the circuit breaker. S202. Use wavelet transform algorithm to extract frequency domain features from waveform data in simulation dataset, use frequency domain features as abstract features of circuit breaker, and use abstract features and physical features together to form a hybrid feature space.

[0032] Wavelet transform algorithms are used to extract frequency domain features from waveform data as abstract features. Wavelet transform offers multi-resolution analysis capabilities, effectively capturing the variation characteristics of waveform data across different frequency ranges. These frequency domain features often contain potential information about circuit breaker faults. For example, certain faults may cause signal enhancement or attenuation at specific frequency components. By using abstract features, this fault information, which is not easily observed directly, can be extracted, complementing the physical features.

[0033] Physical characteristics can intuitively reflect the macroscopic changes in the operating state of a circuit breaker, while abstract characteristics can uncover microscopic fault information hidden in the data. The hybrid feature space constructed by combining the two takes into account both macroscopic and microscopic, direct and indirect feature information, and can more comprehensively describe the state of the circuit breaker. Fault prediction and type determination based on this feature space can significantly improve the accuracy and comprehensiveness of fault identification, and reduce misjudgments or omissions caused by single features.

[0034] S300. A circuit breaker fault prediction model is trained using a neural network algorithm. An additional physical residual loss is designed into the prediction error loss constructed during the training of the neural network algorithm. The prediction error loss and the physical residual loss are combined to form the total loss function. Training samples are extracted from the simulation dataset to perform meta-learning on the fault prediction model. The specific steps for constructing the total loss function by combining the prediction error loss with the physical residual loss are as follows: S301. A circuit breaker fault prediction model is trained using a neural network algorithm. The prediction error loss constructed during the training of the neural network algorithm is L. date The prediction error loss is used to measure the error between the predicted result output by the fault prediction model and the true value; an additional physical residual loss is designed into the prediction error loss constructed during the training of the neural network algorithm, with the following formula: ; In the formula, L physics Represents physical residual loss, Physics law Representing different key physical models, f θ (s) represents the different physical prediction values ​​of the fault prediction model when predicting circuit breaker faults, y physics This indicates that the predicted physical value corresponds to the observed actual physical value. S302. Combine the prediction error loss and the physical residual loss to form the total loss function, as shown in the formula: ; In the formula, L total Let λ represent the total loss and λ represent the hyperparameter. By combining prediction error loss and physical residual loss, and adjusting their weights through the hyperparameter λ, the model's prediction accuracy is guaranteed while adhering to physical laws. This avoids situations where the model violates physical principles in pursuit of high prediction accuracy, making model training more scientific and reasonable.

[0035] S303. Extract different states of the circuit breaker from the simulation dataset, providing k samples for each state to form a support set. Then extract n samples as a query set. Combine the support set and the query set to form a task. Use model-independent meta-learning to learn the support set and the query set. Set up an inner loop and an outer loop. In the inner loop, extract a task and use the initial parameters of the fault prediction model, which represent the initial values ​​of different model parameters in the fault prediction model. Perform forward propagation on the support set, calculate the total loss, calculate the gradient of the total loss with respect to the initial parameters, subtract the gradient from the initial parameters, and repeat the inner loop to obtain the adaptation parameters of the fault prediction model after adapting to the task. In the inner loop, the model is trained on the support set using initial parameters. By calculating the total loss and gradients and updating the parameters, it quickly adapts to the current task. This process enables the model to rapidly adjust its parameters when faced with new tasks (different combinations of states), improving its adaptability to specific tasks.

[0036] In the outer loop, the adapted fault prediction model is used to propagate forward on the query set for each task, and the total loss is calculated. The total loss of the query set for all tasks is summed to form the meta-loss. The gradient of the meta-loss with respect to the initial parameters is calculated, and the optimized parameters are obtained by subtracting the gradient from the initial parameters. In the meta-learning, the outer loop optimizes the initial parameters of the fault prediction model, and the inner loop uses the optimized parameters as the initial parameters to perform an adaptation task to obtain the adapted parameters.

[0037] The outer loop summarizes the total loss of all tasks' query sets to form the meta-loss. By optimizing the initial parameters, the model's initial parameters are made more generalizable, enabling it to quickly adapt to more diverse tasks. The fault prediction model trained through meta-learning can still maintain high prediction accuracy when facing new fault types or new working scenarios that have not been seen in reality, significantly improving the model's generalization ability and reducing its limitations in practical applications.

[0038] S400. For each fault type, feature vectors are extracted in the hybrid feature space and the mean of the feature vectors is calculated to form a "prototype" of the fault type. When the fault prediction model detects a circuit breaker fault, real-time physical data of the circuit breaker is extracted and mapped to the hybrid feature space to calculate the similarity distance with each "prototype" to determine the real-time fault type. The specific steps for determining the type of real-time fault are as follows: S401. In the hybrid feature space, transform the features of each circuit breaker fault type into feature vectors. Calculate the average value of each feature vector for each fault type, and use the average value of each feature vector as the "prototype p" of the corresponding fault type. c ”; S402. When a circuit breaker fault is detected, the circuit breaker data output from the fault prediction model at the time of the fault is mapped into the hybrid feature space. The feature vector of the real-time fault is extracted, and the similarity distance between the real-time fault and each "prototype" of the circuit breaker is calculated. The formula is as follows: ; In the formula, Sim represents the similarity distance between the real-time fault and each "prototype" of the circuit breaker, p(s) test ) represents the feature vector of a real-time fault; the fault type corresponding to the "prototype" with the minimum similarity distance is selected as the real-time fault type.

[0039] When a circuit breaker fault is detected, real-time fault data is mapped to a hybrid feature space to extract feature vectors. The fault type is determined by calculating the similarity distance with each fault "prototype". This method, based on feature vector space distance, intuitively reflects the similarity between the real-time fault and known fault types; the smaller the distance, the higher the similarity. Selecting the fault type corresponding to the minimum similarity distance as the real-time fault type allows for quick and accurate fault type determination, saving valuable time for subsequent fault handling and reducing delays or errors caused by incorrect fault type identification.

[0040] S500: Construct a knowledge graph based on expert experience; search for dimensional actions and causes of real-time fault types in the knowledge graph.

[0041] The specific steps for finding the dimensional actions and causes of real-time fault types in a knowledge graph are as follows: S501. Experts actively input maintenance actions, fault causes, and correlations for different fault types, using fault type, fault cause, fault characteristics, and maintenance actions as nodes and correlations as edges, and construct a knowledge graph based on expert experience. S502. After determining the fault type of a real-time fault, locate the corresponding fault type in the knowledge graph, extract the nodes related to the real-time fault type based on the edge extraction of the knowledge graph, extract the maintenance actions and fault causes of the real-time fault pair, and push them to the maintenance personnel.

[0042] Once the real-time fault type is determined, the corresponding fault type node can be quickly located in the knowledge graph. Related fault causes and maintenance action nodes can then be extracted through the association edges. This query method eliminates the need for manual searching through large amounts of documents or data, enabling the acquisition of accurate and comprehensive fault-related information in a short time. This provides timely decision support for maintenance personnel, helping them quickly develop fault handling plans, improve fault handling efficiency, and reduce the impact of faults on the normal operation of circuit breakers.

[0043] An artificial intelligence-based circuit breaker fault monitoring system includes a data acquisition module, a feature extraction module, a model training module, a fault type lookup module, and a fault reasoning module. The data acquisition module is used to calculate and generate different key physical models using the physical data of the circuit breaker, and to collect physical data on the health status and fault status of the circuit breaker in history. The feature extraction module is used to extract physical features related to the key physical model during circuit breaker operation, extract abstract features from waveform data using signal processing algorithms, and construct a hybrid feature space using physical features and abstract features. The model training module is used to train a circuit breaker fault prediction model using a neural network algorithm, extract training samples from the simulation dataset, and perform meta-learning on the fault prediction model. The fault type lookup module is used to extract real-time physical data of the circuit breaker and map it into a hybrid feature space when the fault prediction model detects a circuit breaker fault, calculate the similarity distance with each "prototype", and determine the real-time fault type. The fault reasoning module is used to construct a knowledge graph based on expert experience; and to search for the dimensional actions and causes of real-time fault types in the knowledge graph.

[0044] The data acquisition module includes key physical model units and simulation units; The key physical model unit is used to calculate and generate different key physical models using the physical data of the circuit breaker. The key physical models include electromagnetic operating mechanism model, mechanical motion model and contact electrical contact model. The simulation unit is used to collect physical data of the historical circuit breaker health status and different fault states, and input them into three models to obtain waveform data to form a simulation dataset.

[0045] The feature extraction module includes physical feature units and abstract feature units; The physical feature unit is used to calculate the correlation coefficient between each type of working data and each type of physical data in the key physical model when the circuit breaker fails, and selects the working data with a correlation coefficient of non-zero as the physical feature related to the key physical model of the circuit breaker. The abstract feature unit is used to extract frequency domain features from waveform data in the simulation dataset using wavelet transform algorithm, and the frequency domain features are used as abstract features of the circuit breaker.

[0046] The model training module includes a total loss function construction unit and a meta-learning unit; The total loss function construction unit is used to combine the prediction error loss and the physical residual loss to form the total loss function; The meta-learning unit is used to extract training samples from the simulation dataset and perform meta-learning on the fault prediction model.

[0047] Example: This embodiment uses a 10kV vacuum circuit breaker as the monitoring object. This type of circuit breaker is widely used in power distribution networks. Common faults include coil breakage, contact wear, and mechanical jamming. If the faults are not detected and dealt with in time, they may lead to serious consequences such as power outages and equipment damage.

[0048] An electromagnetic operating mechanism model was constructed, specifically as follows: Based on the parameters of a 10kV vacuum circuit breaker, the coil DC resistance R = 50Ω and the back electromotive force constant Ke = 0.02V were determined. s / mm, by collecting data on the changes in coil current i and moving iron core displacement x over time t using sensors, and substituting these values ​​into the formula. In this study, the electromagnetic force was measured experimentally under different coil currents (0.3-1.5A) and moving iron core displacements (5-15mm), and the results were obtained through fitting. ; The same collected data were used to construct a mechanical motion model and a contact electrical contact model. The collected physical data (1000 sets each) of four states—historical health status, coil breakage, contact wear, and mechanical jamming—were input into the above three key physical models. The simulation outputs data such as coil current waveform, moving iron core displacement waveform, and contact resistance waveform, resulting in a total of 4000 sets of waveform data, which constitute the simulation dataset.

[0049] Twenty operational data points were collected from the circuit breaker during operation, including coil current, coil voltage, moving iron core displacement, contact pressure, contact resistance, and vibration frequency. When a circuit breaker malfunctioned, the Pearson correlation coefficient was calculated between each operational data point and the physical data in the key physical model (e.g., coil current versus coil current in the electromagnetic operating mechanism model, moving iron core displacement versus moving iron core displacement in the mechanical motion model). Twelve data points with non-zero correlation coefficients (e.g., coil current, moving iron core displacement, contact resistance, and vibration frequency) were selected as physical characteristics.

[0050] Using the wavelet transform tool in Matlab R2021a (with the db4 wavelet basis), multi-resolution decomposition was performed on 4000 sets of waveform data in the simulation dataset. Energy values ​​for each waveform in five frequency bands (10-50Hz, 50-100Hz, 100-200Hz, 200-500Hz, and 500-1000Hz) were extracted, yielding 20 frequency domain features, which were used as abstract features. The 12 physical features were combined with the 20 abstract features to form a hybrid feature space containing 32 feature dimensions. Each sample (waveform data) corresponds to a 32-dimensional feature vector in this space.

[0051] Based on the key physical model, construct the physical residual loss L. physics For example, for the electromagnetic operating mechanism model, calculate the absolute error between the coil voltage U predicted by the model and the actual coil voltage U. Similarly, calculate the prediction residuals of the mechanical motion model and the contact electrical contact model. Set the hyperparameter λ=0.3 and construct the total loss function; A three-layer fully connected neural network was used as the fault prediction model. The input layer consisted of a 32-dimensional feature vector, the hidden layers had 64 and 32 neurons respectively (with ReLU activation function), and the output layer had 4 neurons (corresponding to 4 states, with Softmax activation function). Using the TensorFlow 2.5 framework, 70% (2800 sets) of the simulation dataset was used as the training set, 15% (600 sets) as the validation set, and 15% (600 sets) as the test set. The Adam optimizer (learning rate 0.001) was used to minimize the total loss function, and the model was iteratively trained for 100 epochs to obtain the initial fault prediction model. Model-independent meta-learning (MAML) is used to optimize the initial model. Four states are randomly selected from the simulation dataset. Five samples are selected for each state to form the support set, and three samples are selected to form the query set, forming one task. A total of 200 tasks are generated.

[0052] Then, the fault prediction model was used to predict real-time faults, and the type of real-time fault was determined to be coil breakage. Five experts with over 10 years of experience in circuit breaker maintenance were invited to input fault types, fault causes, maintenance actions, and related relationships. Taking "coil open circuit fault" as an example, fault causes include "coil aging," "loose wiring," and "overvoltage surge," while maintenance actions include "replacing the coil," "checking and tightening the wiring," and "checking the grid voltage and installing a surge protector." A knowledge graph was constructed in the Neo4j graph database, with "coil open circuit fault," "coil aging," and "coil replacement" as nodes and "causing" and "corresponding maintenance" as edges.

[0053] Once the real-time fault type is determined to be coil breakage, the backend system locates the "coil breakage fault" node in the knowledge graph. It extracts fault cause nodes such as "coil aging", "loose wiring" and "overvoltage impact" through the "cause" side, and extracts maintenance action nodes such as "replace coil" and "check wiring and tighten" through the "corresponding maintenance" side. The fault cause and maintenance action are pushed to the mobile APP of the on-site maintenance personnel in text and table form to guide them in carrying out fault handling.

[0054] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A circuit breaker fault monitoring method based on artificial intelligence, characterized in that: The method includes the following steps: S100. Calculate and generate different key physical models using the physical data of the circuit breaker, collect physical data of the health and fault states of the circuit breaker in history, input them into the key physical models for simulation, output waveform data of different key physical models, and form a simulation dataset. S200: Extract physical features related to the key physical model during circuit breaker operation, extract abstract features from waveform data using signal processing algorithms, and construct a hybrid feature space using physical features and abstract features. S300. A circuit breaker fault prediction model is trained using a neural network algorithm. An additional physical residual loss is designed into the prediction error loss constructed during the training of the neural network algorithm. The prediction error loss and the physical residual loss are combined to form the total loss function. Training samples are extracted from the simulation dataset to perform meta-learning on the fault prediction model. S400. For each fault type, feature vectors are extracted in the hybrid feature space and the mean of the feature vectors is calculated to form a "prototype" of the fault type. When the fault prediction model detects a circuit breaker fault, the real-time physical data of the circuit breaker is extracted and mapped to the hybrid feature space to calculate the similarity distance with each "prototype" to determine the real-time fault type. S500: Construct a knowledge graph based on expert experience; search for dimensional actions and causes of real-time fault types in the knowledge graph.

2. The circuit breaker fault monitoring method based on artificial intelligence according to claim 1, characterized in that: The specific steps for forming the simulation dataset in S100 are as follows: S101. The key physical model includes an electromagnetic operating mechanism model, a mechanical motion model, and a contact electrical contact model. The specific construction method of the electromagnetic operating mechanism model is as follows: collect the coil current, coil voltage, electromagnetic force, and moving iron core displacement during circuit breaker operation; and construct the influence relationship between coil current, electromagnetic force, and moving iron core displacement on coil voltage based on the physical principles of circuit breaker operation. The formula is: ； In the formula, U represents the coil voltage, L represents the coil inductance, i represents the coil current, R represents the coil DC resistance, and K represents the coil resistance. e Let x represent the back electromotive force constant, x represent the displacement of the moving iron core, and t represent time. Based on actual conditions, professionals derive the relationship between the coil current and the displacement of the moving iron core on the electromagnetic force as F. magnetic =f(i,x), where F magnetic Let f represent electromagnetic force and f represent the functional relationship; a model of an electromagnetic operating mechanism is constructed using these two influence relationships. S102. The specific method for constructing the mechanical motion model is as follows: Analyze the forces acting on the circuit breaker during operation, collect the spring reaction force, friction force, damping force, and total mass when driving the iron core, and calculate the resultant force of the acceleration generated by the total mass of the circuit breaker. The formula is: ； In the formula, m represents the total mass, and F spring F represents the spring reaction force. friction F represents frictional force. damping Indicates damping force. This represents the second derivative of the displacement of the moving iron core with respect to time. The formula for calculating the resultant force is used as a model of mechanical motion. S103. Professionals actively construct an electrical contact model of the contact based on the relationship between contact resistance and contact pressure, material and degree of burning. The physical data of the historical circuit breaker health status and different fault states are collected and input into three models to obtain waveform data to form a simulation dataset.

3. The circuit breaker fault monitoring method based on artificial intelligence according to claim 2, characterized in that: The specific steps in S200 for constructing a hybrid feature space using physical and abstract features are as follows: S201. Collect all working data involved in the operation of the circuit breaker. When the circuit breaker fails, calculate the correlation coefficient between each type of working data and each type of physical data in the key physical model. Select the working data with a correlation coefficient of non-zero as the physical features related to the key physical model of the circuit breaker. S202. Use wavelet transform algorithm to extract frequency domain features from waveform data in simulation dataset, use frequency domain features as abstract features of circuit breaker, and use abstract features and physical features together to form a hybrid feature space.

4. The circuit breaker fault monitoring method based on artificial intelligence according to claim 3, characterized in that: The specific steps in S300 to construct the total loss function by combining the prediction error loss with the physical residual loss are as follows: S301. A circuit breaker fault prediction model is trained using a neural network algorithm. The prediction error loss constructed during the training of the neural network algorithm is L. date The prediction error loss is used to measure the error between the predicted result output by the fault prediction model and the true value; an additional physical residual loss is designed into the prediction error loss constructed during the training of the neural network algorithm, with the following formula: ； In the formula, L physics Represents physical residual loss, Physics law Representing different key physical models, f θ (s) represents the different physical prediction values ​​of the fault prediction model when predicting circuit breaker faults, y physics This indicates that the predicted physical value corresponds to the observed actual physical value. S302. Combine the prediction error loss and the physical residual loss to form the total loss function, as shown in the formula: ； In the formula, L total Let λ represent the total loss and λ represent the hyperparameter. S303. Extract different states of the circuit breaker from the simulation dataset, providing k samples for each state to form a support set. Then extract n samples as a query set. Combine the support set and the query set to form a task. Use model-independent meta-learning to learn the support set and the query set. Set up an inner loop and an outer loop. In the inner loop, extract a task and use the initial parameters of the fault prediction model, which represent the initial values ​​of different model parameters in the fault prediction model. Perform forward propagation on the support set, calculate the total loss, calculate the gradient of the total loss with respect to the initial parameters, subtract the gradient from the initial parameters, and repeat the inner loop to obtain the adaptation parameters of the fault prediction model after adapting to the task. In the outer loop, the adapted fault prediction model is used to propagate forward in the query set for the task, calculate the total loss, summarize the total loss of the query set of all tasks to form the meta-loss, calculate the gradient of the meta-loss with respect to the initial parameters, and subtract the gradient from the initial parameters to obtain the optimized parameters. In meta-learning, the outer loop optimizes the initial parameters of the fault prediction model, while the inner loop uses the optimized parameters as initial parameters to perform an adaptation task to obtain the adapted parameters.

5. The circuit breaker fault monitoring method based on artificial intelligence according to claim 4, characterized in that: The specific steps for determining the real-time fault type in S400 are as follows: S401. In the hybrid feature space, transform the features of each circuit breaker fault type into feature vectors. Calculate the average value of each feature vector for each fault type, and use the average value of each feature vector as the "prototype p" of the corresponding fault type. c ”; S402. When a circuit breaker fault is detected, the circuit breaker data output from the fault prediction model at the time of the fault is mapped into the hybrid feature space. The feature vector of the real-time fault is extracted, and the similarity distance between the real-time fault and each "prototype" of the circuit breaker is calculated. The formula is as follows: ； In the formula, Sim represents the similarity distance between the real-time fault and each "prototype" of the circuit breaker, p(s) test ) represents the feature vector of a real-time fault; the fault type corresponding to the "prototype" with the minimum similarity distance is selected as the real-time fault type.

6. The circuit breaker fault monitoring method based on artificial intelligence according to claim 5, characterized in that: The specific steps for searching the dimension actions and causes of real-time fault types in the knowledge graph in S500 are as follows: S501. Experts actively input maintenance actions, fault causes, and correlations for different fault types, using fault type, fault cause, fault characteristics, and maintenance actions as nodes and correlations as edges, and construct a knowledge graph based on expert experience. S502. After determining the fault type of a real-time fault, locate the corresponding fault type in the knowledge graph, extract the nodes related to the real-time fault type based on the edge extraction of the knowledge graph, extract the maintenance actions and fault causes of the real-time fault pair, and push them to the maintenance personnel.

7. A circuit breaker fault monitoring system based on artificial intelligence, characterized in that: The circuit breaker fault monitoring system includes a data acquisition module, a feature extraction module, a model training module, a fault type lookup module, and a fault reasoning module. The data acquisition module is used to calculate and generate different key physical models using the physical data of the circuit breaker, and to collect physical data on the health status and fault status of the circuit breaker in history. The feature extraction module is used to extract physical features related to the key physical model during circuit breaker operation, extract abstract features from waveform data using signal processing algorithms, and construct a hybrid feature space using physical features and abstract features. The model training module is used to train a circuit breaker fault prediction model using a neural network algorithm, extract training samples from the simulation dataset, and perform meta-learning on the fault prediction model. The fault type lookup module is used to extract real-time physical data of the circuit breaker and map it into a hybrid feature space when the fault prediction model detects a circuit breaker fault, to calculate the similarity distance with each "prototype" and determine the real-time fault type. The fault reasoning module is used to construct a knowledge graph based on expert experience; and to search for the dimensional actions and causes of real-time fault types in the knowledge graph.

8. The circuit breaker fault monitoring system based on artificial intelligence according to claim 7, characterized in that: The data acquisition module includes a key physical model unit and a simulation unit; The key physical model unit is used to calculate and generate different key physical models using the physical data of the circuit breaker. The key physical models include electromagnetic operating mechanism model, mechanical motion model and contact electrical contact model. The simulation unit is used to collect physical data of the historical circuit breaker health status and different fault states, and input them into three models to obtain waveform data to form a simulation dataset.

9. The circuit breaker fault monitoring system based on artificial intelligence according to claim 7, characterized in that: The feature extraction module includes physical feature units and abstract feature units; The physical feature unit is used to calculate the correlation coefficient between each type of working data and each type of physical data in the key physical model when the circuit breaker fails, and selects the working data with a correlation coefficient of non-zero as the physical feature related to the key physical model of the circuit breaker. The abstract feature unit is used to extract frequency domain features from waveform data in the simulation dataset using wavelet transform algorithm, and the frequency domain features are used as abstract features of the circuit breaker.

10. A circuit breaker fault monitoring system based on artificial intelligence according to claim 7, characterized in that: The model training module includes a total loss function construction unit and a meta-learning unit; The total loss function construction unit is used to combine the prediction error loss and the physical residual loss to form the total loss function; The meta-learning unit is used to extract training samples from the simulation dataset and perform meta-learning on the fault prediction model.

Images