A classification method and system for electromagnetic interference signal of partial discharge live detection in traction substation

Abstract

The present invention relates to the technical field of electromagnetic interference signal processing, and provides a classification method and system for electromagnetic interference signals of live detection of partial discharge in traction substations, so as to solve the problems that the prior art is difficult to accurately classify various types of electromagnetic interference signals, difficult to adapt to changes under different working conditions, poor adaptive adjustment ability, and insufficient adaptability. The present invention includes: constructing a dynamic graph model to process electromagnetic interference signal data streams and generate fused data; performing adaptive situational awareness classification on data in real time by combining reinforcement learning and deep belief networks; extracting the spatiotemporal spectrum characteristics of electromagnetic interference signals, constructing an interactive expert system, generating target parameter configurations, and obtaining preliminary classification results; adjusting various algorithm parameters to generate final classification results. The technical solution provided by the present invention provides an efficient and intelligent electromagnetic interference signal processing method, which ensures the accuracy and reliability of the classification results and improves the efficiency and effectiveness of electromagnetic environment management.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of electromagnetic interference signal processing, and in particular to a classification method and system for electromagnetic interference signal of live detection of partial discharge in a traction substation. Background Art

[0002] Traction substation is an important part of the power system, and its stable operation is directly related to the safety and reliability of the entire power grid. Partial discharge live detection is one of the key means to evaluate the health status of substation equipment. However, in the actual detection process, electromagnetic interference signals are often mixed with partial discharge signals, which seriously affects the accuracy and reliability of detection. Therefore, efficient and accurate classification of electromagnetic interference signals has become a key technical requirement to ensure the effectiveness of partial discharge live detection.

[0003] At present, the detection of electromagnetic interference signals mainly relies on traditional spectrum analysis methods, such as fast Fourier transform, which are difficult to achieve accurate classification of various types of electromagnetic interference signals. In addition, existing solutions mostly use fixed thresholds or rules to determine abnormal situations, which are difficult to adapt to changes under different working conditions; existing methods have poor adaptive adjustment capabilities and insufficient adaptability. Summary of the invention

[0004] The embodiments of the present invention provide a classification method and system for electromagnetic interference signals of live detection of partial discharge in a traction substation, so as to solve the problems in the prior art of difficulty in accurately classifying various types of electromagnetic interference signals, difficulty in adapting to changes in different working conditions, poor adaptive adjustment capability and insufficient adaptability.

[0005] In a first aspect, an embodiment of the present invention provides a classification method for electromagnetic interference signals of live detection of partial discharge in a traction substation, comprising:

[0006] Collect electromagnetic interference signal data streams, build a dynamic graph model using a graph neural network, use the dynamic graph model to process the electromagnetic interference signal data stream to obtain fused data, and use an isolation forest algorithm to perform real-time anomaly detection processing on the fused data to obtain an abnormal interference signal;

[0007] Combining a reinforcement learning algorithm and a deep belief network, performing adaptive situational awareness classification processing on the abnormal interference signal to obtain a classification result, and identifying different types of electromagnetic interference signals based on the classification result;

[0008] Combining convolutional neural network and recurrent neural network to extract the spatiotemporal spectrum characteristics of the electromagnetic interference signal;

[0009] Constructing an interactive expert system based on a Bayesian network and a genetic algorithm, using the interactive expert system to analyze the spatiotemporal spectrum characteristics to obtain a system analysis result, quantitatively evaluating the importance of different types of electromagnetic interference signals based on the system analysis result to obtain an importance evaluation result, generating a target parameter configuration corresponding to each type of electromagnetic interference signal based on the importance evaluation result and existing domain knowledge, and generating a preliminary classification result based on the target parameter configuration, the analysis result and the domain knowledge;

[0010] The parameters of each algorithm are adjusted using a fast feedback path and a slow feedback path to obtain adjusted parameters, and the preliminary classification results are optimized based on the adjusted parameters and the domain knowledge to obtain final classification results. The algorithms include the reinforcement learning algorithm, the deep belief network, the convolutional neural network, and the recurrent neural network.

[0011] Optionally, the interactive expert system is used to analyze the spatiotemporal spectrum characteristics to obtain a system analysis result, and based on the system analysis result, the importance of different types of electromagnetic interference signals is quantitatively evaluated to obtain an importance evaluation result, and based on the importance evaluation result and existing domain knowledge, a target parameter configuration corresponding to each type of electromagnetic interference signal is generated, including:

[0012] The interactive expert system is used to analyze the spatiotemporal spectrum characteristics, and the probability of occurrence of different types of electromagnetic interference signals and the degree of influence of the electromagnetic interference signals on the partial discharge live detection process are obtained by updating the probability distribution of each node in the Bayesian network, wherein the nodes are different types of electromagnetic interference signals;

[0013] Based on the occurrence probability and the impact degree, the importance of the electromagnetic interference signal is quantitatively evaluated to generate a target importance evaluation result for each type of electromagnetic interference signal, and the genetic algorithm is used to adjust the weight value of each type of electromagnetic interference signal during the quantitative evaluation process to optimize the importance evaluation result;

[0014] Based on the importance evaluation result and existing domain knowledge, the genetic algorithm is used to search for the best parameter configuration corresponding to each electromagnetic interference signal type in a predefined parameter space, and after the user adjusts the best parameter configuration, the target parameter configuration is obtained.

[0015] Optionally, based on the occurrence probability and the impact degree, the importance of the electromagnetic interference signal is quantitatively evaluated to generate a target importance evaluation result of each type of electromagnetic interference signal, and the genetic algorithm is applied in the quantitative evaluation process to adjust the weight value of each type of electromagnetic interference signal to optimize the importance evaluation result, including:

[0016] Based on the occurrence probability and the impact degree, a quantitative evaluation model is constructed, and the importance score of each type of electromagnetic interference signal is calculated by the evaluation function in the quantitative evaluation model to generate a preliminary importance evaluation result of each type of electromagnetic interference signal;

[0017] In the quantitative evaluation process, a genetic algorithm is applied to iteratively optimize the weight value of each type of electromagnetic interference signal through population initialization, selection operation, crossover operation, mutation operation, and evaluation and update operation until a preset stop condition is met to obtain an optimized weight value, wherein the stop condition is that the performance index of the quantitative evaluation model does not change after the weight value is applied;

[0018] Based on the optimized weight values, the preliminary importance evaluation results are optimized to generate target importance evaluation results for each type of electromagnetic interference signal.

[0019] Optionally, the parameters of each algorithm are adjusted by using a fast feedback path and a slow feedback path to obtain adjusted parameters, and the preliminary classification result is optimized based on the adjusted parameters and the domain knowledge to obtain a final classification result, wherein each algorithm includes the reinforcement learning algorithm, the deep belief network, the convolutional neural network and the recurrent neural network, including:

[0020] Based on the instant detection results and user operation feedback information collected in real time by the fast feedback path, the parameters of the reinforcement learning algorithm and the deep belief network are adjusted by using an online gradient descent method to obtain an adjusted first parameter;

[0021] Based on the system operation data collected in real time by the slow feedback path, the performance deviation of the convolutional neural network and the recurrent neural network is evaluated by using the Markov chain Monte Carlo method to obtain a performance deviation result, and based on the performance deviation result, the hyperparameters of the convolutional neural network and the recurrent neural network are optimized to generate adjusted second parameters;

[0022] Based on the adjusted first parameter, the adjusted second parameter and the domain knowledge, the preliminary classification result is optimized to generate a final classification result.

[0023] Optionally, the instant detection results and user operation feedback information collected in real time based on the fast feedback path are used to adjust the parameters of the reinforcement learning algorithm and the deep belief network using an online gradient descent method to obtain the adjusted first parameters, including:

[0024] Based on a fast feedback path, the electromagnetic interference signal strength, the spatiotemporal spectrum characteristics and the abnormal interference signal are collected in real time to obtain an immediate detection result;

[0025] Based on the fast feedback path, parameter adjustment suggestions proposed by users are collected in real time, and users' daily operation behaviors are monitored to obtain user operation feedback information, wherein the daily operation behaviors include switch device operation and detection sensitivity adjustment operation;

[0026] Based on the instant detection result and the user operation feedback information, a loss function is constructed, and the gradient value of the loss function relative to the reinforcement learning algorithm and the deep belief network is calculated through a back propagation algorithm. The parameters of the reinforcement learning algorithm and the deep belief network are updated in real time based on the gradient value to obtain the adjusted first parameter.

[0027] Optionally, the system operation data collected in real time based on the slow feedback path is evaluated using a Markov chain Monte Carlo method to obtain a performance deviation result of the convolutional neural network and the recurrent neural network, and based on the performance deviation result, the hyperparameters of the convolutional neural network and the recurrent neural network are optimized to generate adjusted second parameters, including:

[0028] Based on a slow feedback path, collecting system operation data in real time, the system operation data including output results of the convolutional neural network and the recurrent neural network, the output results including the spatiotemporal spectrum characteristics and the time domain characteristics of the electromagnetic interference signal;

[0029] Based on the system operation data, a Markov chain Monte Carlo method is used to randomly select multiple target sample configurations from a predefined probability distribution, and the processing effect of the convolutional neural network and the recurrent neural network on the electromagnetic interference signal under each target sample configuration is simulated and evaluated to obtain multiple simulation evaluation results, and the multiple simulation evaluation results are counted and analyzed to obtain performance deviation results;

[0030] Based on the performance deviation result, hyperparameters to be optimized are determined, the hyperparameters to be optimized are combined into a parameter space, and a Bayesian optimization technique is used to search for an optimal solution in the parameter space, where the optimal solution represents the adjusted second parameter.

[0031] Optionally, the combination of the reinforcement learning algorithm and the deep belief network to perform adaptive situational awareness classification processing on the abnormal interference signal to obtain a classification result includes:

[0032] Using a deep belief network to perform feature extraction processing on the abnormal interference signal, obtaining feature codes representing the original signal pattern and deep-level features, constructing a reinforcement learning environment containing different types of electromagnetic interference signals according to the feature codes, and designing a reward mechanism, and generating a learning framework based on the reinforcement learning environment and the reward mechanism;

[0033] Based on the learning framework, a reinforcement learning algorithm is applied to perform adaptive classification processing, and the existing classification strategy is optimized by updating the existing classification strategy and situational awareness to obtain an optimized classification strategy. In the situational awareness process, a situational representation is constructed, and a situational recognition result is obtained based on the situational representation. Based on the situational recognition result, a Bayesian network is used to analyze the probability dependency relationship between different influencing factors to obtain a probability dependency relationship analysis result. A situational model is constructed based on the probability dependency relationship analysis result. Based on the situational model and predefined situational triggering rules, the existing classification strategy is adjusted. The classification strategy is a method for identifying and distinguishing different types of electromagnetic interference signals. The influencing factors include the characteristics of the electromagnetic interference signal, external environmental conditions, and device status.

[0034] Based on the optimized classification strategy and the feature coding, the abnormal interference signal is comprehensively classified to generate a classification result.

[0035] In a second aspect, an embodiment of the present invention provides a classification system for detecting electromagnetic interference signals during partial discharge in a traction substation, including:

[0036] A collection module is used to collect electromagnetic interference signal data streams, build a dynamic graph model using a graph neural network, process the electromagnetic interference signal data stream using the dynamic graph model to obtain fused data, and perform real-time anomaly detection processing on the fused data using an isolation forest algorithm to obtain an abnormal interference signal;

[0037] An identification module is used to combine a reinforcement learning algorithm and a deep belief network to perform adaptive situational awareness classification processing on the abnormal interference signal to obtain a classification result, and identify different types of electromagnetic interference signals based on the classification result;

[0038] An extraction module, used to combine a convolutional neural network and a recurrent neural network to extract the spatiotemporal spectrum characteristics of the electromagnetic interference signal;

[0039] A construction module is used to construct an interactive expert system based on a Bayesian network and a genetic algorithm, use the interactive expert system to analyze the spatiotemporal spectrum characteristics to obtain a system analysis result, quantitatively evaluate the importance of different types of electromagnetic interference signals based on the system analysis result to obtain an importance evaluation result, generate a target parameter configuration corresponding to each type of electromagnetic interference signal based on the importance evaluation result and existing domain knowledge, and generate a preliminary classification result based on the target parameter configuration, the analysis result and the domain knowledge;

[0040] An adjustment module is used to adjust the parameters of each algorithm using a fast feedback path and a slow feedback path to obtain adjusted parameters, and optimize the preliminary classification results based on the adjusted parameters and the domain knowledge to obtain final classification results. The algorithms include the reinforcement learning algorithm, the deep belief network, the convolutional neural network, and the recurrent neural network.

[0041] In a third aspect, an embodiment of the present invention provides a computing device, comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a method for grading electromagnetic interference signals for live detection of partial discharge in a traction substation as described in any one of the first aspects.

[0042] In a fourth aspect, an embodiment of the present invention provides a computer storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement a method for grading electromagnetic interference signals for live detection of partial discharge in a traction substation as described in any one of the first aspects.

[0043] Beneficial effects of the present invention: In the embodiment of the present invention, an electromagnetic interference signal data stream is collected, a dynamic graph model is constructed using a graph neural network, the electromagnetic interference signal data stream is processed using the dynamic graph model to obtain fused data, and the fused data is processed in real time using an isolation forest algorithm to obtain an abnormal interference signal; a reinforcement learning algorithm and a deep belief network are combined to perform adaptive context-aware classification processing on the abnormal interference signal to obtain a classification result, and different types of electromagnetic interference signals are identified based on the classification result; a convolutional neural network and a recurrent neural network are combined to extract the spatiotemporal spectral characteristics of the electromagnetic interference signal; an interactive expert system based on a Bayesian network and a genetic algorithm is constructed, and the interactive expert system is used to classify the electromagnetic interference signal. The method comprises the following steps: analyzing the spatiotemporal spectrum characteristics to obtain a system analysis result, quantitatively evaluating the importance of different types of electromagnetic interference signals based on the system analysis result to obtain an importance evaluation result, generating a target parameter configuration corresponding to each type of electromagnetic interference signal based on the importance evaluation result and existing domain knowledge, and generating a preliminary classification result based on the target parameter configuration, the analysis result and the domain knowledge; adjusting the parameters of each algorithm using a fast feedback path and a slow feedback path to obtain adjusted parameters, optimizing the preliminary classification result based on the adjusted parameters and the domain knowledge to obtain a final classification result, wherein each algorithm includes a reinforcement learning algorithm, a deep belief network, a convolutional neural network and a recurrent neural network.

[0044] (1) The technical solution provided by the present invention provides an efficient and intelligent electromagnetic interference signal processing method, which ensures the accuracy and reliability of the classification results and improves the efficiency and effectiveness of electromagnetic environment management. Among them, by introducing a quantitative evaluation model and a genetic algorithm, the accuracy of the importance evaluation of different types of electromagnetic interference signals is significantly improved, and the risk of misjudgment is reduced; the target importance evaluation results generated based on the optimized weight value ensure the rationality of the evaluation and provide reliable guidance for subsequent processing.

[0045] (2) The present invention combines a fast feedback path, an online gradient descent method, a slow feedback path, and a Markov chain Monte Carlo method to adjust various algorithm parameters in real time, significantly improving the accuracy and response speed of detection, and ensuring that the system remains efficient and stable in long-term operation; combining the adjusted parameters and domain knowledge to optimize the preliminary classification results, generate more accurate and reliable final classification results, and reduce false positives and negatives; the intelligent feedback mechanism and parameter optimization not only enhance the practicality and reliability of the system, but also provide solid decision-making support for maintenance and management.

[0046] These and other aspects of the present invention will become more apparent from the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 A flow chart of a classification method for electromagnetic interference signal of live detection of partial discharge in a traction substation provided by an embodiment of the present invention;

[0049] Figure 2 A schematic diagram of the structure of a classification system for live detection of electromagnetic interference signals of partial discharge in a traction substation provided by an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of the structure of a computing device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0052] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0054] Figure 1 A flowchart of a classification method for detecting electromagnetic interference signals of partial discharge live in a traction substation is provided in accordance with an embodiment of the present invention. Figure 1 As shown, the method includes:

[0055] In the detection of partial discharge live electricity in traction substations, the presence of electromagnetic interference signals seriously affects the accuracy and reliability of the detection results. Traditional detection methods are difficult to effectively distinguish and classify complex electromagnetic interference signals, resulting in frequent false alarms and missed alarms. In order to meet this challenge and improve the intelligence and adaptability of the detection system, based on this, the present invention provides a classification method for electromagnetic interference signals for partial discharge live electricity detection in traction substations, such as Figure 1 ,include:

[0056] Step 101: Collect electromagnetic interference signal data streams, build a dynamic graph model using a graph neural network, use the dynamic graph model to process the electromagnetic interference signal data stream to obtain fused data, and use an isolation forest algorithm to perform real-time anomaly detection on the fused data to obtain an abnormal interference signal;

[0057] In this step, the electromagnetic interference signal data stream refers to the electromagnetic interference signal continuously collected from the traction substation partial discharge live detection equipment, usually including time domain and frequency domain information. Graph Neural Network (GNN) refers to a neural network model used to process graph structured data, which can capture the complex relationship between nodes. Dynamic graph model refers to a model built based on graph neural network, which can update the relationship between nodes and edges over time to adapt to changing data streams. Fusion data refers to the comprehensive data processed by the dynamic graph model, which contains the spatiotemporal characteristics of the original electromagnetic interference signal and their mutual relationship. Isolation Forest Algorithm refers to an unsupervised learning algorithm used for real-time anomaly detection, especially suitable for outlier detection in high-dimensional data.

[0058] This step collects electromagnetic interference signals in the traction substation through the sensor network. These signals include multi-dimensional information such as voltage, current, and frequency, forming a continuous data stream, providing comprehensive and real-time data support for subsequent processing; the collected electromagnetic interference signals are converted into graph structure data, where each node represents a sensor or signal source, and the edge represents the correlation between nodes. The dynamic graph model is constructed through the graph neural network, which can update the relationship between nodes and edges in real time. This step aims to capture the complex spatiotemporal associations between electromagnetic interference signals and improve the feature expression ability; the electromagnetic interference signal data stream is processed through the dynamic graph model, multi-level spatiotemporal features are extracted, and these features are fused into fused data. This step aims to generate fused data containing rich spatiotemporal information and provide high-quality input for subsequent processing; the isolation forest algorithm is applied to perform real-time anomaly detection on the fused data, identify abnormal points that do not conform to the normal pattern, and mark them as abnormal interference signals. This step aims to quickly screen out potential abnormal signals and reduce false positives and false negatives.

[0059] Step 102: combining a reinforcement learning algorithm and a deep belief network, performing adaptive situational awareness classification processing on the abnormal interference signal to obtain a classification result, and identifying different types of electromagnetic interference signals based on the classification result;

[0060] In this step, reinforcement learning (RL) refers to a machine learning method that continuously optimizes strategies by interacting with the environment to maximize cumulative rewards. Deep belief network (DBN) refers to a deep generative model that is good at extracting high-level features of complex data. Adaptive context-aware classification processing refers to a method that combines reinforcement learning algorithms and deep belief networks to automatically adjust classification strategies according to the current situation and improve classification accuracy.

[0061] In this step, a deep belief network is used to extract features of abnormal interference signals and generate high-level feature codes. Based on these codes, a reinforcement learning environment containing different types of electromagnetic interference signals is constructed, and a reward mechanism is designed to generate a learning framework. Based on the learning framework, a reinforcement learning algorithm is applied for adaptive classification processing. By updating the existing classification strategy and optimizing the classification strategy with contextual awareness, efficient classification of different types of electromagnetic interference signals is achieved based on the optimized classification strategy, ensuring the accuracy and adaptability of the classification results.

[0062] Step 103: Combining a convolutional neural network and a recurrent neural network to extract the spatiotemporal spectrum characteristics of the electromagnetic interference signal;

[0063] In this step, Convolutional Neural Network (CNN) refers to a neural network commonly used in image recognition and is also suitable for extracting spectral features. Recurrent Neural Network (RNN) refers to a neural network that is good at processing sequence data and is suitable for extracting time series features. Spatiotemporal spectrum features refer to information containing time, space, and spectrum dimensions, which are used to fully describe the characteristics of electromagnetic interference signals.

[0064] In this step, a convolutional neural network is used to extract the spectral features of the electromagnetic interference signal, and a recurrent neural network is used to extract its time series features. The two are combined to generate spatiotemporal spectral features. This step aims to fully capture the time and frequency domain characteristics of the electromagnetic interference signal and provide detailed information for subsequent analysis.

[0065] Step 104: construct an interactive expert system based on a Bayesian network and a genetic algorithm, use the interactive expert system to analyze the spatiotemporal spectrum characteristics to obtain a system analysis result, quantitatively evaluate the importance of different types of electromagnetic interference signals based on the system analysis result to obtain an importance evaluation result, generate a target parameter configuration corresponding to each type of electromagnetic interference signal based on the importance evaluation result and existing domain knowledge, and generate a preliminary classification result based on the target parameter configuration, the analysis result and the domain knowledge;

[0066] In this step, Bayesian Network refers to a probabilistic graphical model used to represent the conditional dependencies between variables. Genetic Algorithm (GA) refers to a heuristic search algorithm that simulates natural selection and genetic mechanisms to find the optimal solution. Interactive Expert System refers to a system that combines Bayesian Network and Genetic Algorithm to make intelligent decisions using domain knowledge.

[0067] This step builds an interactive expert system based on Bayesian networks, combines genetic algorithms to optimize parameters, analyzes spatiotemporal spectrum characteristics, and generates system analysis results. This step uses intelligent analysis tools to comprehensively evaluate electromagnetic interference signals and improve the scientificity and accuracy of decision-making; based on the system analysis results, calculate the importance score of each type of electromagnetic interference signal and generate importance assessment results. This step clarifies the degree of influence of various types of electromagnetic interference signals and provides priority guidance for subsequent processing; combines the importance assessment results and the knowledge of domain experts to generate the optimal parameter configuration for each type of electromagnetic interference signal. This step ensures the rationality and effectiveness of the parameter configuration and improves processing efficiency; comprehensively considers the target parameter configuration, analysis results and domain knowledge to generate preliminary classification results, which provide a basis for the final classification results.

[0068] Step 105: adjusting the parameters of each algorithm using a fast feedback path and a slow feedback path to obtain adjusted parameters, optimizing the preliminary classification result based on the adjusted parameters and the domain knowledge to obtain a final classification result, wherein the algorithms include the reinforcement learning algorithm, the deep belief network, the convolutional neural network, and the recurrent neural network;

[0069] This step collects instant detection results and user operation feedback information in real time through a fast feedback path, and uses the online gradient descent method to adjust the parameters of the reinforcement learning algorithm and the deep belief network; it collects system operation data in real time through a slow feedback path, and uses the Markov chain Monte Carlo method to evaluate the performance deviation of the convolutional neural network and the recurrent neural network, and optimizes the hyperparameters. This step ensures that the parameters of each algorithm are always in the optimal state, thereby improving the stability and adaptability of the system; the preliminary grading results are optimized in combination with the adjusted parameters and domain knowledge to generate the final grading results. This step ensures the accuracy and reliability of the final grading results.

[0070] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0071] By introducing a variety of advanced algorithms and technologies, the classification accuracy of different types of electromagnetic interference signals has been significantly improved, reducing false positives and false negatives.

[0072] Through adaptive situational awareness and classification processing, the system can maintain stable performance in complex and changing environments and adapt to different working conditions.

[0073] Through the combination of convolutional neural network and recurrent neural network, the temporal and spatial spectrum characteristics of electromagnetic interference signals are fully extracted, which improves the integrity and accuracy of feature expression.

[0074] The fast feedback path and the slow feedback path are used to adjust the parameters of each algorithm in real time, ensuring the optimization of parameter configuration and improving the overall performance of the system;

[0075] By building an interactive expert system and combining domain knowledge to quantitatively evaluate the importance of electromagnetic interference signals, a scientific basis is provided for maintenance decisions, enhancing the practicality and reliability of the system.

[0076] After the electromagnetic interference signal is preliminarily classified and feature extracted, in order to further improve the accuracy and reliability of the analysis, based on this, the present invention provides a specific embodiment, the step 104, constructing a dynamic data fusion framework, the dynamic data fusion framework adjusts the association rules between the structured data and the unstructured data according to the real-time analysis requirements, specifically including the following steps:

[0077] Step 401: Analyze the spatiotemporal spectrum characteristics by using the interactive expert system, and obtain the occurrence probability of different types of electromagnetic interference signals and the degree of influence of the electromagnetic interference signals on the partial discharge live detection process by updating the probability distribution of each node in the Bayesian network, wherein the nodes are different types of electromagnetic interference signals;

[0078] In this step, the node refers to a variable or type in the Bayesian network, which refers to different types of electromagnetic interference signals in this embodiment. The probability of occurrence refers to the possibility of a certain type of electromagnetic interference signal appearing in a specific situation. The degree of influence refers to the specific impact of a certain type of electromagnetic interference signal on the partial discharge live detection process, such as false alarm rate, missed alarm rate, etc.

[0079] This step uses the Bayesian network module in the interactive expert system to input the spatiotemporal spectrum features. The Bayesian network updates the probability distribution of each node within it based on these features. Each node represents a type of electromagnetic interference signal, and its probability distribution reflects the possibility of the occurrence of this type of signal in the current situation and its impact on the detection of partial discharge live electricity. This step accurately identifies the key characteristics of various types of electromagnetic interference signals and evaluates their impact on the detection process, providing a basis for subsequent classification.

[0080] Step 402: Based on the occurrence probability and the impact degree, the importance of the electromagnetic interference signal is quantitatively evaluated to generate a target importance evaluation result for each type of electromagnetic interference signal. During the quantitative evaluation process, the genetic algorithm is used to adjust the weight value of each type of electromagnetic interference signal to optimize the importance evaluation result.

[0081] In this step, the weight value refers to the different importance ratios given to different types of electromagnetic interference signals in the quantitative evaluation.

[0082] In this step, a quantitative evaluation model is constructed to calculate the importance score of each type of electromagnetic interference signal, and a genetic algorithm is applied to adjust the weight value of each type of electromagnetic interference signal to optimize the importance score and ensure that the importance score of each type of electromagnetic interference signal is more accurate and reasonable.

[0083] Step 403: Based on the importance evaluation result and the existing domain knowledge, the genetic algorithm is used to search for the best parameter configuration corresponding to each electromagnetic interference signal type in the predefined parameter space, and after the user adjusts the best parameter configuration, the target parameter configuration is obtained;

[0084] This step defines the parameter space, including the hyperparameters of various processing algorithms, such as learning rate, regularization coefficient, etc.; uses a genetic algorithm to search for the best configuration in this parameter space and maximize the evaluation function, which comprehensively considers the importance evaluation results and the knowledge of domain experts; submits the best searched parameter configuration to the user for adjustment, and finally determines the target parameter configuration.

[0085] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0086] By introducing Bayesian networks and genetic algorithms, the accuracy of importance assessment of different types of electromagnetic interference signals is significantly improved, reducing the risk of misjudgment.

[0087] The application of genetic algorithms enables the system to automatically adjust weights and parameter configurations in changing situations, improving the system's adaptive capabilities;

[0088] Target parameter configuration ensures that the parameters of each processing algorithm are always in the optimal state, improving the overall performance;

[0089] Through the intelligent analysis of interactive expert systems, reliable decision support is provided for maintenance and management, enhancing the practicality and reliability of the system.

[0090] The present invention provides a specific embodiment, wherein the step 402 performs a quantitative evaluation on the importance of the electromagnetic interference signal based on the occurrence probability and the impact degree, generates a target importance evaluation result of each type of electromagnetic interference signal, and applies the genetic algorithm to adjust the weight value of each type of electromagnetic interference signal during the quantitative evaluation process to optimize the importance evaluation result, specifically comprising the following steps:

[0091] Step 301: constructing a quantitative evaluation model based on the occurrence probability and the impact degree, calculating the importance score of each type of electromagnetic interference signal through the evaluation function in the quantitative evaluation model, and generating a preliminary importance evaluation result of each type of electromagnetic interference signal;

[0092] This step builds a quantitative evaluation model based on the probability of occurrence and the degree of impact. By defining an evaluation function to calculate the importance score of each type of electromagnetic interference signal, the importance of each type of electromagnetic interference signal is preliminarily quantified, and a preliminary importance evaluation result of each type of electromagnetic interference signal is generated, providing a basis for subsequent optimization.

[0093] More specifically, an embodiment of the present invention provides an importance scoring formula, which introduces a Gaussian mixture model to more accurately capture the probability distribution of different types of electromagnetic interference signals, and combines a multi-objective optimization function to perform importance score calculation. The specific calculation formula is as follows:

[0094] ;

[0095] in, Indicates Importance score of electromagnetic interference-like signals; Represents the number of components of the Gaussian mixture model; Indicates The weights of the components satisfy ; Indicates The next Probability of occurrence of electromagnetic interference-like signals; Indicates The next The degree of influence of electromagnetic interference-like signals on discharge detection results; The weight coefficient representing the probability of occurrence; The weight coefficient indicating the degree of influence satisfies ; is the entropy function used to measure uncertainty; is the weight coefficient of the entropy function;

[0096] The importance scoring formula significantly improves the accuracy and reliability of the evaluation of different types of electromagnetic interference signals by introducing the Gaussian mixture model and entropy function. The formula introduces the Gaussian mixture model to more accurately capture the probability distribution of different types of electromagnetic interference signals, thereby improving the accuracy of classification; by combining the probability of occurrence and the degree of impact, and introducing the entropy function to measure uncertainty, the quantitative evaluation model can comprehensively consider multiple factors and improve the reliability of the importance score; by setting the weight coefficient and , the weight of different factors in the score can be flexibly adjusted to meet the needs of different application scenarios.

[0097] Among them, the specific calculation formula of the entropy function is as follows:

[0098] ;

[0099] in, is the entropy function used to measure uncertainty; Indicates The probability of occurrence of quasi-electromagnetic interference signals under different conditions; Indicates The impact degree of electromagnetic interference signals under different conditions; For all possible states Sum all possible states Indicates the specific performance status of each type of electromagnetic interference signal under different conditions or situations;

[0100] Among them, the specific calculation formula of the weight coefficient of the entropy function is as follows:

[0101] ;

[0102] ;

[0103] in, is the weight coefficient of the entropy function; Indicates The uncertainty measure of the electromagnetic interference-like signal in the current situation is defined as the standard deviation of the probability of occurrence and the degree of impact; Represents the maximum uncertainty measure in all types of electromagnetic interference signals, used for normalization , ensuring that its value range is within between; is the variance of the probability of occurrence; is the variance of the effect.

[0104] The weight coefficient formula of the entropy function introduces standard deviation as an uncertainty measure, which can reflect the volatility and complexity of electromagnetic interference signals in the current situation. The larger the standard deviation, the more unstable the probability of occurrence and impact of such signals, and a higher weight should be given to measure their uncertainty; by dividing by ,Will Normalized to [0,1] to ensure that its value is always within a reasonable range. This can avoid the weight imbalance problem caused by the large difference in absolute values ​​in different situations, making the model more stable and reliable. This formula allows the quantitative evaluation model to adaptively adjust the weight of the entropy function according to the actual situation. When the uncertainty of a certain type of electromagnetic interference signal is high, the quantitative evaluation model will automatically assign a larger value, thereby paying more attention to the uncertainty assessment of this type of signal. Conversely, it effectively improves the flexibility and adaptability of the model.

[0105] Step 302: applying a genetic algorithm in the quantitative evaluation process, iteratively optimizing the weight value of each type of electromagnetic interference signal through population initialization, selection operation, crossover operation, mutation operation, and evaluation and update operation, until a preset stop condition is met, and obtaining an optimized weight value, wherein the stop condition is that the performance index of the quantitative evaluation model does not change after applying the weight value;

[0106] This step uses a genetic algorithm to randomly generate a set of initial weight values, each weight value combination constitutes an individual, forming an initial population; select excellent individuals to enter the next generation population according to the fitness function; perform crossover operations on the selected individuals to generate new individuals; perform mutation operations on the new individuals with a certain probability to ensure population diversity; evaluate the fitness of the new individuals and replace the individuals with lower fitness until the preset stopping conditions are met, such as the maximum number of iterations or the performance indicators of the quantitative evaluation model do not change after applying the weight values.

[0107] Step 303: Optimizing the preliminary importance evaluation result based on the optimized weight value to generate a target importance evaluation result for each type of electromagnetic interference signal;

[0108] This step uses the optimized weight value to recalculate the importance score of each type of electromagnetic interference signal to generate the final target importance evaluation result. By comparing the importance scores before and after optimization, it is ensured that the optimized result better meets the actual needs.

[0109] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0110] By introducing quantitative evaluation models and genetic algorithms, the accuracy of importance evaluation of different types of electromagnetic interference signals is significantly improved, reducing the risk of misjudgment.

[0111] The application of genetic algorithms enables the system to automatically adjust weight values ​​in changing situations, improving the system's adaptive capabilities;

[0112] The target importance evaluation results generated based on the optimized weight values ​​ensure the rationality of the importance evaluation and provide reliable guidance for subsequent processing;

[0113] Through intelligent evaluation and optimization, reliable decision support is provided for maintenance and management, enhancing the practicality and reliability of the system.

[0114] The present invention provides a specific embodiment, in which the step 105 uses a fast feedback path and a slow feedback path to adjust the parameters of each algorithm to obtain adjusted parameters, and optimizes the preliminary classification result based on the adjusted parameters and the domain knowledge to obtain a final classification result. The algorithms include the reinforcement learning algorithm, the deep belief network, the convolutional neural network, and the recurrent neural network, and specifically include the following steps:

[0115] Step 501: Based on the instant detection results and user operation feedback information collected in real time by the fast feedback path, the parameters of the reinforcement learning algorithm and the deep belief network are adjusted by using an online gradient descent method to obtain adjusted first parameters;

[0116] In this step, the fast feedback path refers to the real-time collection of instant detection results and user operation feedback information, supporting online parameter adjustment. The slow feedback path refers to the real-time collection of system operation data, supporting offline performance evaluation and hyperparameter optimization. The instant detection result refers to the electromagnetic interference signal detection result obtained in real time from the detection equipment. User operation feedback information refers to the operation instructions and feedback provided by the user during use. Online Gradient Descent (OGD) refers to a machine learning optimization algorithm that can update model parameters in real time based on new data.

[0117] This step collects instant detection results and user operation feedback information in real time through the sensor network and user interface; the collected data is input into the reinforcement learning algorithm and deep belief network, and the parameters of these models are updated in real time using the online gradient descent method. Each update is based on the current batch of data to ensure that the model can quickly adapt to new changes; after multiple iterative updates, the optimized reinforcement learning algorithm and deep belief network parameter settings are obtained. For example, in a traction substation, a high-frequency noise is detected in real time, and the user confirms that the noise is a false alarm through the interface. The system immediately collects this feedback information and adjusts the parameters of the reinforcement learning algorithm and deep belief network through the online gradient descent method.

[0118] Step 502: Based on the system operation data collected in real time by the slow feedback path, the performance deviation of the convolutional neural network and the recurrent neural network is evaluated using the Markov chain Monte Carlo method to obtain a performance deviation result, and based on the performance deviation result, the hyperparameters of the convolutional neural network and the recurrent neural network are optimized to generate adjusted second parameters;

[0119] In this step, Markov Chain Monte Carlo (MCMC) refers to a statistical simulation method used to evaluate model performance deviations and perform parameter optimization.

[0120] This step regularly collects system operation data, including long-term series of electromagnetic interference signals, detection accuracy, false alarm rate, etc.; the Markov chain Monte Carlo method is used to evaluate the performance of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to generate performance deviation results. This method simulates the model performance under different parameter combinations through random sampling to find the best parameter settings. According to the performance deviation results, the hyperparameters of CNN and RNN, such as learning rate, regularization coefficient, batch size, etc., are adjusted to generate the adjusted second parameters. For example, the collected data show that CNN has low detection accuracy in certain frequency ranges and RNN has insufficient prediction ability on long sequences. The MCMC method is used to evaluate these deviations, and it is found that the learning rate of CNN and the regularization coefficient of RNN are the main influencing factors. After multiple rounds of sampling and evaluation, the optimal learning rate and regularization coefficient are finally determined to generate the adjusted second parameters. These optimized parameters significantly improve the detection accuracy and stability of CNN and RNN.

[0121] Step 503: Based on the adjusted first parameter, the adjusted second parameter and the domain knowledge, the preliminary classification result is optimized to generate a final classification result;

[0122] This step combines the adjusted first parameter and the adjusted second parameter to form a comprehensive parameter configuration, and combines the knowledge of domain experts, such as known electromagnetic interference signal characteristics, common problems and solutions, to further optimize the preliminary classification results. Based on the comprehensive parameter configuration and domain knowledge, the importance score of each type of electromagnetic interference signal is recalculated to generate the final classification results. For example, the preliminary classification results show that a certain type of high-frequency noise is misjudged as a high-priority signal. Combined with the adjusted first and second parameters, the importance of this type of signal is re-evaluated. At the same time, domain experts pointed out that this type of noise usually does not have a significant impact on partial discharge detection, so its importance score is adjusted from 0.8 to 0.5. The final generated classification results show that this type of high-frequency noise is correctly classified as a low-priority signal, reducing unnecessary alarms.

[0123] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0124] Through fast feedback paths and online gradient descent, the parameters of the reinforcement learning algorithm and deep belief network are adjusted in real time, which significantly improves the accuracy and response speed of detection;

[0125] Through slow feedback paths and Markov chain Monte Carlo methods, the performance of convolutional neural networks and recurrent neural networks is regularly evaluated and optimized to ensure that the system remains efficient and stable in long-term operation;

[0126] Combine the adjusted parameters and domain knowledge to optimize the preliminary classification results, generate more accurate and reliable final classification results, and reduce false positives and false negatives;

[0127] Through intelligent feedback mechanism and parameter optimization, reliable decision support is provided for maintenance and management, enhancing the practicality and reliability of the system.

[0128] The present invention provides a specific embodiment, wherein step 501, based on the instant detection results and user operation feedback information collected in real time by the fast feedback path, uses an online gradient descent method to adjust the parameters of the reinforcement learning algorithm and the deep belief network to obtain the adjusted first parameters, specifically includes the following steps:

[0129] Step 511: Based on the fast feedback path, the electromagnetic interference signal strength, the spatiotemporal spectrum characteristics and the abnormal interference signal are collected in real time to obtain an instant detection result;

[0130] This step is based on a fast feedback path to collect the electromagnetic interference signal strength and spatiotemporal spectrum characteristics in the traction substation in real time, and use the isolation forest algorithm to detect abnormal interference signals. The above information is integrated into instant detection results and provided to subsequent processing modules.

[0131] Step 512: Based on the fast feedback path, parameter adjustment suggestions proposed by the user are collected in real time, and the user's daily operation behavior is monitored to obtain user operation feedback information, wherein the daily operation behavior includes switch device operation and detection sensitivity adjustment operation;

[0132] In this step, users can propose adjustment suggestions for model parameters through the interface, such as improving the detection sensitivity of a certain type of signal; the system monitors the user's daily operation behavior in real time, such as switch device operation and detection sensitivity adjustment operation, and records the user's operation habits and preferences; the parameter adjustment suggestions and daily operation behaviors proposed by the user are integrated into user operation feedback information. For example, the user suggests increasing the detection sensitivity of high-frequency noise through the interface, and frequently adjusts the detection sensitivity in daily operation. The system records this information and forms user operation feedback information, that is, the parameter adjustment suggestion is to improve the detection sensitivity of high-frequency noise, and the daily operation behavior is to frequently adjust the detection sensitivity.

[0133] Step 513: constructing a loss function based on the instant detection result and the user operation feedback information, calculating the gradient value of the loss function relative to the reinforcement learning algorithm and the deep belief network through a back propagation algorithm, and updating the parameters of the reinforcement learning algorithm and the deep belief network in real time based on the gradient value to obtain an adjusted first parameter;

[0134] In this step, the loss function refers to a function used to measure the difference between the model's predicted value and the actual value, guiding model optimization. The backpropagation algorithm refers to a commonly used neural network training algorithm that updates parameters by calculating the gradient of the loss function relative to the model parameters. The gradient value refers to the rate of change of the loss function relative to the model parameters, which is used to guide the direction of parameter update.

[0135] This step constructs a loss function based on the instant detection results and user operation feedback information to measure the difference between the model prediction value and the actual value. The back propagation algorithm is used to calculate the gradient value of the loss function relative to the RL and DBN parameters; based on the gradient value, the online gradient descent method is used to update the RL and DBN parameters in real time to obtain the adjusted first parameter. For example, the instant detection results show that high-frequency noise is misjudged as a high-priority signal, and the user recommends improving the detection sensitivity; the loss function is constructed, and it is found that the model has a large error in high-frequency noise detection; the back propagation algorithm is used to calculate the gradient value, and it is found that the RL reward function parameters and the number of hidden layer nodes of DBN have a significant impact on the loss; after multiple iterative updates, the optimal parameters are finally determined, and the adjusted first parameters are obtained.

[0136] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0137] By collecting key information of electromagnetic interference signals and user operation feedback information in real time, building a loss function and updating model parameters in real time, the detection accuracy and response speed are significantly improved;

[0138] Combine the parameter adjustment suggestions and daily operation behaviors proposed by users to enable the system to better meet user needs and improve user experience;

[0139] The back-propagation algorithm is used to calculate the gradient value and update the parameters in real time, ensuring that the reinforcement learning algorithm and the deep belief network are always in the optimal state, thus improving the overall performance of the model.

[0140] Through intelligent feedback mechanism and parameter optimization, reliable decision support is provided for maintenance and management, enhancing the practicality and reliability of the system.

[0141] The present invention provides a specific embodiment, wherein step 502, based on the system operation data collected in real time by the slow feedback path, uses the Markov chain Monte Carlo method to evaluate the performance deviation of the convolutional neural network and the recurrent neural network, obtains a performance deviation result, and optimizes the hyperparameters of the convolutional neural network and the recurrent neural network based on the performance deviation result to generate an adjusted second parameter, specifically comprising the following steps:

[0142] Step 521: Based on the slow feedback path, collecting system operation data in real time, the system operation data including the output results of the convolutional neural network and the recurrent neural network, the output results including the spatiotemporal spectrum characteristics and the time domain characteristics of the electromagnetic interference signal;

[0143] This step is based on the slow feedback path to collect system operation data in real time to ensure that the data is representative. The system operation data includes the output results of CNN and RNN, especially the spatiotemporal spectrum characteristics and time domain characteristics of electromagnetic interference signals. These data reflect the performance of the model in actual application scenarios. For example, at 2 a.m. every day, the system automatically collects the operation data of the previous day. Assume that the data of a certain day is as follows: the spatiotemporal spectrum characteristics are that the high-frequency noise is concentrated between 500MHz and 800MHz, and periodic pulses appear every 5 seconds; the time domain characteristics are that the maximum peak of the high-frequency noise is -20dBm, and the average power is -30dBm; the peak of the periodic pulse is -15dBm, and the duration is 1 millisecond.

[0144] Step 522: Based on the system operation data, a Markov chain Monte Carlo method is used to randomly select multiple target sample configurations from a predefined probability distribution, and a simulation evaluation is performed on the processing effect of the convolutional neural network and the recurrent neural network on the electromagnetic interference signal under each target sample configuration to obtain multiple simulation evaluation results, and the multiple simulation evaluation results are counted and analyzed to obtain a performance deviation result;

[0145] In this step, the target sample configuration refers to multiple parameter combinations randomly selected from a predefined probability distribution for simulation evaluation.

[0146] This step randomly selects multiple hyperparameter combinations from a predefined probability distribution based on system operation data to form multiple target sample configurations; simulates and evaluates the processing effects of CNN and RNN under each target sample configuration to generate multiple simulation evaluation results. Evaluation indicators may include detection accuracy, false alarm rate, false negative rate, etc.; statistics and analysis are performed on multiple simulation evaluation results, and statistical quantities such as average and standard deviation are calculated to identify the performance deviation results of the model under different configurations.

[0147] Step 523: Based on the performance deviation result, determine the hyperparameters to be optimized, combine the hyperparameters to be optimized into a parameter space, and use Bayesian optimization technology to search for an optimal solution in the parameter space, where the optimal solution represents the adjusted second parameter;

[0148] In this step, Bayesian optimization technology refers to an efficient global optimization algorithm that finds the optimal solution by building a proxy model. The optimal solution refers to the best hyperparameter combination found in the parameter space, which represents the adjusted second parameter.

[0149] In this step, based on the performance deviation results, key hyperparameters that need to be optimized, such as learning rate and regularization coefficient, are determined; the key hyperparameters to be optimized are combined into a parameter space and their value ranges are defined. For example, learning rate η∈[0.0001,0.01], regularization coefficient λ∈[0.00001,0.001]; the Bayesian optimization technique is used to search for the best solution in the parameter space. This technique gradually approaches the global optimal solution by constructing a proxy model, such as Gaussian process regression; the optimal hyperparameter combination finally found represents the adjusted second parameter.

[0150] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0151] By regularly collecting system operation data and using MCMC methods for simulation evaluation, we identified the performance differences of the model under different configurations, thereby optimizing key hyperparameters and significantly improving the detection accuracy and stability of CNN and RNN.

[0152] Bayesian optimization technology can efficiently search for the best solution in complex parameter space, ensuring that the model can maintain high performance under various working conditions and enhancing the robustness of the system;

[0153] The automated optimization process reduces the workload of manual parameter adjustment and improves the system's adaptive capabilities;

[0154] Through intelligent evaluation and optimization, reliable decision support is provided for maintenance and management, enhancing the practicality and reliability of the system.

[0155] The present invention provides a specific embodiment, wherein step 102 combines a reinforcement learning algorithm and a deep belief network to perform adaptive context-aware classification processing on the abnormal interference signal to obtain a classification result, specifically comprising the following steps:

[0156] Step 201: using a deep belief network to perform feature extraction processing on the abnormal interference signal to obtain feature codes representing the original signal pattern and deep-level features, constructing a reinforcement learning environment containing different types of electromagnetic interference signals according to the feature codes, and designing a reward mechanism, and generating a learning framework based on the reinforcement learning environment and the reward mechanism;

[0157] In this step, feature encoding refers to the encoding of the original signal pattern and deep features extracted by the deep belief network for subsequent processing. Reinforcement learning environment refers to the simulation of actual application scenarios, including different types of electromagnetic interference signals, and the design of reward mechanism to guide strategy optimization. Reward mechanism refers to the incentive rules defined in the reinforcement learning environment, which is used to evaluate and feedback the quality of model behavior. Learning framework refers to the overall learning architecture built on the reinforcement learning environment and reward mechanism, which is used to train and optimize classification strategies.

[0158] In this step, a deep belief network is used to extract features from abnormal interference signals and generate feature codes. These codes not only contain the pattern information of the original signal, but also extract deep features, such as spectrum distribution, time series characteristics, etc. Based on the feature codes, a simulation environment containing multiple types of electromagnetic interference signals is constructed. Each signal type corresponds to a different feature code; a reward mechanism is defined, such as positive rewards for correctly classifying high-priority signals and negative rewards for false alarms or omissions; the reinforcement learning environment and the reward mechanism are combined to form a complete learning framework for training and optimizing classification strategies. For example, a section of high-frequency noise is detected, and its spatiotemporal spectrum characteristics show the presence of periodic pulses. The feature codes are extracted through a deep belief network to generate the following features: the spectrum distribution is concentrated between 500MHz and 800MHz; the time series characteristics are a peak every 5 seconds with a duration of 1 millisecond; other deep features are power spectral density, autocorrelation function, etc.

[0159] Step 202: Based on the learning framework, a reinforcement learning algorithm is applied to perform adaptive classification processing, and the existing classification strategy is optimized by updating the existing classification strategy and situational awareness to obtain an optimized classification strategy. In the situational awareness process, a situational representation is constructed, and a situational recognition result is obtained based on the situational representation. Based on the situational recognition result, a Bayesian network is used to analyze the probability dependency relationship between different influencing factors to obtain a probability dependency relationship analysis result. A situational model is constructed based on the probability dependency relationship analysis result. Based on the situational model and predefined situational triggering rules, the existing classification strategy is adjusted. The classification strategy is a method for identifying and distinguishing different types of electromagnetic interference signals. The influencing factors include the characteristics of the electromagnetic interference signal, external environmental conditions, and device status.

[0160] In this step, situational awareness refers to identifying the current environmental state by constructing situational representation and adjusting the classification strategy according to the situation. The situational representation refers to the description of the current environmental state, including the characteristics of the electromagnetic interference signal, external environmental conditions, and device status. The situational recognition result refers to the determination of the current situation category by analyzing the situational representation. The situational model refers to the model constructed based on the results of the probabilistic dependency analysis, which is used to guide situational awareness and strategy adjustment.

[0161] This step uses reinforcement learning algorithms, such as Q-learning or DQN, to perform adaptive classification processing in the learning framework. The existing classification strategy is continuously updated by updating the existing classification strategy and situational awareness to obtain the optimized classification strategy. Contextual awareness builds a situational representation, identifies the current environmental state, and adjusts the classification strategy based on the situational recognition results. Combined with the situational awareness results, the Bayesian network is used to analyze the probabilistic dependency between different influencing factors, build a situational model, and adjust the classification strategy according to the situational triggering rules. For example, in the simulation environment, the initial classification strategy can correctly classify 80% of high-frequency noise and 70% of periodic pulses. The classification strategy is continuously updated by the reinforcement learning algorithm, and it is found that in certain situations, such as when the humidity is high, periodic pulses are more likely to be misjudged as high-frequency noise. The situational representation is built to identify the current high humidity situation, and the Bayesian network is used to analyze the probabilistic dependency between humidity, signal strength and frequency distribution to obtain the analysis results. The situational model is built based on the analysis results, and the classification strategy is adjusted according to the situational triggering rules. Finally, the optimized classification strategy can correctly classify 95% of high-frequency noise and 90% of periodic pulses.

[0162] Step 203: Based on the optimized classification strategy and the feature coding, the abnormal interference signal is subjected to comprehensive classification processing to generate a classification result.

[0163] This step combines the optimized classification strategy and feature coding to finally classify the abnormal interference signal; output the classification results of each type of electromagnetic interference signal and provide it to the subsequent processing module. For example, a newly detected abnormal interference signal is classified, and the feature coding shows that it has mixed characteristics of high-frequency noise and periodic pulses; using the optimized classification strategy, the system determines that the signal is closer to a periodic pulse, and the final classification result is a periodic pulse.

[0164] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0165] The deep belief network is used to extract deep features and combined with the reinforcement learning algorithm for adaptive classification processing, which significantly improves the accuracy and stability of classification.

[0166] The introduction of situational awareness and Bayesian network analysis enables the system to automatically adjust the classification strategy in different environments, thus enhancing the system's adaptive ability;

[0167] By building situational models and situational triggering rules, the system can dynamically adjust the classification strategy according to the current environmental status and provide more reliable decision support;

[0168] The optimized classification strategy can better identify and distinguish different types of electromagnetic interference signals, reduce false positives and false negatives, and improve system reliability;

[0169] The intelligent classification and processing process reduces the need for manual intervention and improves the user's operating experience and work efficiency.

[0170] Figure 2 The present invention provides a structural schematic diagram of a classification system for detecting electromagnetic interference signals of partial discharge in a traction substation, such as Figure 2 As shown, the system includes:

[0171] A collection module 21 is used to collect electromagnetic interference signal data streams, construct a dynamic graph model using a graph neural network, process the electromagnetic interference signal data stream using the dynamic graph model to obtain fused data, and perform real-time anomaly detection processing on the fused data using an isolation forest algorithm to obtain an abnormal interference signal;

[0172] The identification module 22 is used to combine the reinforcement learning algorithm and the deep belief network to perform adaptive situational awareness classification processing on the abnormal interference signal to obtain a classification result, and identify different types of electromagnetic interference signals based on the classification result;

[0173] An extraction module 23, used to extract the spatiotemporal spectrum characteristics of the electromagnetic interference signal by combining a convolutional neural network and a recurrent neural network;

[0174] A construction module 24 is used to construct an interactive expert system based on a Bayesian network and a genetic algorithm, using the interactive expert system to analyze the spatiotemporal spectrum characteristics to obtain a system analysis result, quantitatively evaluate the importance of different types of electromagnetic interference signals based on the system analysis result to obtain an importance evaluation result, generate a target parameter configuration corresponding to each type of electromagnetic interference signal based on the importance evaluation result and existing domain knowledge, and generate a preliminary classification result based on the target parameter configuration, the analysis result and the domain knowledge;

[0175] The adjustment module 25 is used to adjust the parameters of each algorithm using a fast feedback path and a slow feedback path to obtain adjusted parameters, and optimize the preliminary classification results based on the adjusted parameters and the domain knowledge to obtain final classification results. The algorithms include the reinforcement learning algorithm, the deep belief network, the convolutional neural network, and the recurrent neural network.

[0176] Figure 2 The grading system for detecting electromagnetic interference signals of partial discharge in traction substation can be implemented Figure 1The implementation principle and technical effect of the classification method of electromagnetic interference signal for partial discharge detection in traction substation described in the embodiment are not described in detail. The specific way in which each module and unit performs the operation of the classification system of electromagnetic interference signal for partial discharge detection in traction substation in the above embodiment has been described in detail in the embodiment of the method, and will not be described in detail here.

[0177] In one possible design, Figure 2 The grading system for detecting electromagnetic interference signals of partial discharge in a traction substation in the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0178] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0179] The processing component 32 is used to: collect electromagnetic interference signal data streams, build a dynamic graph model using a graph neural network, use the dynamic graph model to process the electromagnetic interference signal data stream to obtain fused data, use the isolation forest algorithm to perform real-time anomaly detection processing on the fused data to obtain abnormal interference signals; combine the reinforcement learning algorithm and the deep belief network to perform adaptive situational awareness classification processing on the abnormal interference signal to obtain classification results, and identify different types of electromagnetic interference signals based on the classification results; combine the convolutional neural network and the recurrent neural network to extract the spatiotemporal spectrum characteristics of the electromagnetic interference signal; build an interactive expert system based on the Bayesian network and the genetic algorithm, and use the interactive expert system to analyze the spatiotemporal spectrum characteristics of the electromagnetic interference signal. characteristics, obtain system analysis results, quantitatively evaluate the importance of different types of electromagnetic interference signals based on the system analysis results to obtain importance evaluation results, generate target parameter configurations corresponding to each type of electromagnetic interference signal based on the importance evaluation results and existing domain knowledge, and generate preliminary classification results based on the target parameter configurations, the analysis results and the domain knowledge; use fast feedback paths and slow feedback paths to adjust the parameters of each algorithm to obtain adjusted parameters, optimize the preliminary classification results based on the adjusted parameters and the domain knowledge to obtain final classification results, and the algorithms include the reinforcement learning algorithm, the deep belief network, the convolutional neural network and the recurrent neural network.

[0180] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0181] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0182] Of course, the computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0183] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc.

[0184] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0185] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0186] The embodiment of the present invention further provides a computer storage medium storing a computer program, which can achieve the above-mentioned Figure 1 The illustrated embodiment is a classification method for electromagnetic interference signals of live detection of partial discharge in a traction substation.

[0187] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0188] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0189] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A classification method for electromagnetic interference signals of partial discharge live detection in traction substations, characterized in that: include: Collect electromagnetic interference signal data streams, build a dynamic graph model using a graph neural network, use the dynamic graph model to process the electromagnetic interference signal data stream to obtain fused data, and use an isolation forest algorithm to perform real-time anomaly detection processing on the fused data to obtain an abnormal interference signal; Combining a reinforcement learning algorithm and a deep belief network, performing adaptive situational awareness classification processing on the abnormal interference signal to obtain a classification result, and identifying different types of electromagnetic interference signals based on the classification result; Combining convolutional neural network and recurrent neural network to extract the spatiotemporal spectrum characteristics of the electromagnetic interference signal; Constructing an interactive expert system based on a Bayesian network and a genetic algorithm, using the interactive expert system to analyze the spatiotemporal spectrum characteristics to obtain a system analysis result, quantitatively evaluating the importance of different types of electromagnetic interference signals based on the system analysis result to obtain an importance evaluation result, generating a target parameter configuration corresponding to each type of electromagnetic interference signal based on the importance evaluation result and existing domain knowledge, and generating a preliminary classification result based on the target parameter configuration, the analysis result and the domain knowledge; The parameters of each algorithm are adjusted using a fast feedback path and a slow feedback path to obtain adjusted parameters, and the preliminary classification results are optimized based on the adjusted parameters and the domain knowledge to obtain final classification results. The algorithms include the reinforcement learning algorithm, the deep belief network, the convolutional neural network, and the recurrent neural network.

2. A classification method for electromagnetic interference signals of partial discharge live detection in traction substation according to claim 1, characterized in that: The method comprises: analyzing the spatiotemporal spectrum characteristics by using the interactive expert system to obtain a system analysis result; quantitatively evaluating the importance of different types of electromagnetic interference signals based on the system analysis result to obtain an importance evaluation result; and generating a target parameter configuration corresponding to each type of electromagnetic interference signal based on the importance evaluation result and existing domain knowledge, including: The interactive expert system is used to analyze the spatiotemporal spectrum characteristics, and the probability of occurrence of different types of electromagnetic interference signals and the degree of influence of the electromagnetic interference signals on the partial discharge live detection process are obtained by updating the probability distribution of each node in the Bayesian network, wherein the nodes are different types of electromagnetic interference signals; Based on the occurrence probability and the impact degree, the importance of the electromagnetic interference signal is quantitatively evaluated to generate a target importance evaluation result for each type of electromagnetic interference signal, and the genetic algorithm is used to adjust the weight value of each type of electromagnetic interference signal during the quantitative evaluation process to optimize the importance evaluation result; Based on the importance evaluation result and existing domain knowledge, the genetic algorithm is used to search for the best parameter configuration corresponding to each electromagnetic interference signal type in a predefined parameter space, and after the user adjusts the best parameter configuration, the target parameter configuration is obtained.

3. A classification method for electromagnetic interference signals of partial discharge live detection in traction substation according to claim 2, characterized in that: The importance of the electromagnetic interference signal is quantitatively evaluated based on the occurrence probability and the impact degree to generate a target importance evaluation result of each type of electromagnetic interference signal, and the weight value of each type of electromagnetic interference signal is adjusted by using the genetic algorithm during the quantitative evaluation process to optimize the importance evaluation result, including: Based on the occurrence probability and the impact degree, a quantitative evaluation model is constructed, and the importance score of each type of electromagnetic interference signal is calculated by the evaluation function in the quantitative evaluation model to generate a preliminary importance evaluation result of each type of electromagnetic interference signal; In the quantitative evaluation process, a genetic algorithm is applied to iteratively optimize the weight value of each type of electromagnetic interference signal through population initialization, selection operation, crossover operation, mutation operation, and evaluation and update operation until a preset stop condition is met to obtain an optimized weight value, wherein the stop condition is that the performance index of the quantitative evaluation model does not change after the weight value is applied; Based on the optimized weight values, the preliminary importance evaluation results are optimized to generate target importance evaluation results for each type of electromagnetic interference signal.

4. A classification method for electromagnetic interference signals of partial discharge live detection in traction substation according to claim 1, characterized in that: The method comprises adjusting the parameters of each algorithm by using the fast feedback path and the slow feedback path to obtain the adjusted parameters, optimizing the preliminary classification result based on the adjusted parameters and the domain knowledge to obtain the final classification result, wherein the algorithms include the reinforcement learning algorithm, the deep belief network, the convolutional neural network and the recurrent neural network, including: Based on the instant detection results and user operation feedback information collected in real time by the fast feedback path, the parameters of the reinforcement learning algorithm and the deep belief network are adjusted by using an online gradient descent method to obtain an adjusted first parameter; Based on the system operation data collected in real time by the slow feedback path, the performance deviation of the convolutional neural network and the recurrent neural network is evaluated by using the Markov chain Monte Carlo method to obtain a performance deviation result, and based on the performance deviation result, the hyperparameters of the convolutional neural network and the recurrent neural network are optimized to generate adjusted second parameters; Based on the adjusted first parameter, the adjusted second parameter and the domain knowledge, the preliminary classification result is optimized to generate a final classification result.

5. A classification method for electromagnetic interference signals of partial discharge live detection in traction substation according to claim 4, characterized in that: The instant detection results and user operation feedback information collected in real time based on the fast feedback path are adjusted using an online gradient descent method to adjust the parameters of the reinforcement learning algorithm and the deep belief network to obtain the adjusted first parameters, including: Based on a fast feedback path, the electromagnetic interference signal strength, the spatiotemporal spectrum characteristics and the abnormal interference signal are collected in real time to obtain an immediate detection result; Based on the fast feedback path, parameter adjustment suggestions proposed by users are collected in real time, and users' daily operation behaviors are monitored to obtain user operation feedback information, wherein the daily operation behaviors include switch device operation and detection sensitivity adjustment operation; Based on the instant detection result and the user operation feedback information, a loss function is constructed, and the gradient value of the loss function relative to the reinforcement learning algorithm and the deep belief network is calculated through a back propagation algorithm. The parameters of the reinforcement learning algorithm and the deep belief network are updated in real time based on the gradient value to obtain the adjusted first parameter.

6. A classification method for electromagnetic interference signals of partial discharge live detection in traction substation according to claim 4, characterized in that: The system operation data collected in real time based on the slow feedback path is evaluated using the Markov chain Monte Carlo method to obtain a performance deviation result of the convolutional neural network and the recurrent neural network, and based on the performance deviation result, the hyperparameters of the convolutional neural network and the recurrent neural network are optimized to generate an adjusted second parameter, including: Based on a slow feedback path, collecting system operation data in real time, the system operation data including output results of the convolutional neural network and the recurrent neural network, the output results including the spatiotemporal spectrum characteristics and the time domain characteristics of the electromagnetic interference signal; Based on the system operation data, a Markov chain Monte Carlo method is used to randomly select multiple target sample configurations from a predefined probability distribution, and the processing effect of the convolutional neural network and the recurrent neural network on the electromagnetic interference signal under each target sample configuration is simulated and evaluated to obtain multiple simulation evaluation results, and the multiple simulation evaluation results are counted and analyzed to obtain performance deviation results; Based on the performance deviation result, hyperparameters to be optimized are determined, the hyperparameters to be optimized are combined into a parameter space, and a Bayesian optimization technique is used to search for an optimal solution in the parameter space, where the optimal solution represents the adjusted second parameter.

7. A classification method for electromagnetic interference signals of partial discharge live detection in traction substation according to claim 1, characterized in that: The combination of the reinforcement learning algorithm and the deep belief network performs adaptive situational awareness classification processing on the abnormal interference signal to obtain a classification result, including: Using a deep belief network to perform feature extraction processing on the abnormal interference signal, obtaining feature codes representing the original signal pattern and deep-level features, constructing a reinforcement learning environment containing different types of electromagnetic interference signals according to the feature codes, and designing a reward mechanism, and generating a learning framework based on the reinforcement learning environment and the reward mechanism; Based on the learning framework, a reinforcement learning algorithm is applied to perform adaptive classification processing, and the existing classification strategy is optimized by updating the existing classification strategy and situational awareness to obtain an optimized classification strategy. In the situational awareness process, a situational representation is constructed, and a situational recognition result is obtained based on the situational representation. Based on the situational recognition result, a Bayesian network is used to analyze the probability dependency relationship between different influencing factors to obtain a probability dependency relationship analysis result. A situational model is constructed based on the probability dependency relationship analysis result. Based on the situational model and predefined situational triggering rules, the existing classification strategy is adjusted. The classification strategy is a method for identifying and distinguishing different types of electromagnetic interference signals. The influencing factors include the characteristics of the electromagnetic interference signal, external environmental conditions, and device status. Based on the optimized classification strategy and the feature coding, the abnormal interference signal is comprehensively classified to generate a classification result.

8. A classification system for detecting electromagnetic interference signals of partial discharge in traction substations, characterized in that: include: A collection module is used to collect electromagnetic interference signal data streams, build a dynamic graph model using a graph neural network, process the electromagnetic interference signal data stream using the dynamic graph model to obtain fused data, and perform real-time anomaly detection processing on the fused data using an isolation forest algorithm to obtain an abnormal interference signal; An identification module is used to combine a reinforcement learning algorithm and a deep belief network to perform adaptive situational awareness classification processing on the abnormal interference signal to obtain a classification result, and identify different types of electromagnetic interference signals based on the classification result; An extraction module, used to combine a convolutional neural network and a recurrent neural network to extract the spatiotemporal spectrum characteristics of the electromagnetic interference signal; A construction module is used to construct an interactive expert system based on a Bayesian network and a genetic algorithm, use the interactive expert system to analyze the spatiotemporal spectrum characteristics to obtain a system analysis result, quantitatively evaluate the importance of different types of electromagnetic interference signals based on the system analysis result to obtain an importance evaluation result, generate a target parameter configuration corresponding to each type of electromagnetic interference signal based on the importance evaluation result and existing domain knowledge, and generate a preliminary classification result based on the target parameter configuration, the analysis result and the domain knowledge; An adjustment module is used to adjust the parameters of each algorithm using a fast feedback path and a slow feedback path to obtain adjusted parameters, and optimize the preliminary classification results based on the adjusted parameters and the domain knowledge to obtain final classification results. The algorithms include the reinforcement learning algorithm, the deep belief network, the convolutional neural network, and the recurrent neural network.

9. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a classification method for electromagnetic interference signals of live detection of partial discharge in a traction substation as described in any one of claims 1-7.

10. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, a classification method for electromagnetic interference signals of live detection of partial discharge in a traction substation as described in any one of claims 1 to 7 is implemented.

Images