Power communication network fault prediction method, system and device based on deep learning

Abstract

The present invention discloses a method, system and device for predicting faults in a power communication network based on deep learning, and relates to the technical field of predicting faults in a power communication network. The method for predicting faults in a power communication network based on deep learning includes: integrating and processing the acquired historical electric variables and historical magnetic variables to obtain a historical electromagnetic comprehensive data set; training and verifying a power communication network fault prediction model based on the historical electromagnetic comprehensive data set, wherein the power communication network fault prediction model outputs the probability of a fault in the power communication network; predicting a real-time electromagnetic comprehensive data set composed of real-time electric variables and real-time magnetic variables based on the trained and verified power communication network fault prediction model to obtain the real-time fault probability of the power communication network; if the real-time fault probability of the power communication network is greater than or equal to the set fault threshold probability, then obtaining the fault type of the power communication network based on the analysis of the real-time electromagnetic comprehensive data set, and through the combination of electric variables and magnetic variables, the power communication network fault prediction method can provide more comprehensive, more efficient and more accurate fault diagnosis and prediction capabilities.

Description

Technical Field

[0001] The present invention relates to the technical field of power communication network fault prediction, and specifically to a power communication network fault prediction method, system and device based on deep learning. Background Art

[0002] The power distribution network refers to a power network that receives electric energy from the power transmission network or regional power plants, and distributes it locally or step by step according to voltage to various users through distribution facilities. It is composed of overhead lines, cables, towers, distribution transformers, disconnectors, reactive power compensators and some ancillary facilities, and plays an important role in distributing electric energy in the power grid. During the operation of the power distribution network, different faults may occur under different environmental conditions. At present, the fault can only be eliminated after it occurs. When eliminating the fault, the power distribution network needs to be shut down, which affects normal use.

[0003] A Chinese patent application with publication number CN113358977A discloses a method for predicting faults in a distribution power communication network, comprising the following steps: S1, setting a distribution power communication network for training; S2, adjusting the fault type and corresponding environmental conditions of the distribution power communication network for training; S3, inputting the characteristics of the adjusted fault type into the training model, and marking the fault type with the characteristic as fault F; S4, constructing a deep learning model; S5, connecting the deep learning model to the distribution power communication network to be detected, and using the deep learning model to detect the fault condition of the distribution power communication network; S6, collecting statistics on the fault condition of the distribution power communication network and the environmental conditions when the fault occurs; S7, predicting the fault condition of the distribution power communication network according to the environmental conditions.

[0004] However, the existing technology is difficult to combine multiple variables for fault prediction, and there are problems of low fault prediction efficiency and low accuracy. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a power communication network fault prediction method, system and device based on deep learning. Through the combination of electrical variables and magnetic variables, the power communication network fault prediction method can provide more comprehensive, efficient and accurate fault diagnosis and prediction capabilities.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for predicting power communication network faults based on deep learning, comprising the following steps: integrating and processing the acquired historical electrical variables and historical magnetic variables to obtain a historical electromagnetic comprehensive data set; training and verifying a power communication network fault prediction model based on the historical electromagnetic comprehensive data set, wherein the output of the power communication network fault prediction model is the probability of a power communication network failure; predicting a real-time electromagnetic comprehensive data set composed of real-time electrical variables and real-time magnetic variables based on the trained and verified power communication network fault prediction model to obtain a real-time fault probability of the power communication network; if the real-time fault probability of the power communication network is greater than or equal to a set fault threshold probability, obtaining the power communication network fault type based on analysis of the real-time electromagnetic comprehensive data set.

[0007] Furthermore, the historical electromagnetic comprehensive data set includes basic characteristics of historical electrical variables, basic characteristics of historical magnetic variables, cross-characteristics of historical electromagnetic variables, historical electrical periodic characteristics, historical electrical non-periodic characteristics, historical magnetic periodic characteristics and historical magnetic non-periodic characteristics.

[0008] Furthermore, the acquired historical electrical variables and historical magnetic variables are processed and integrated to obtain a historical electromagnetic comprehensive data set, including: unifying the timestamps and data formats of the acquired historical electrical variables and historical magnetic variables; performing basic feature processing on the historical electrical variables and historical magnetic variables respectively to obtain basic features of the historical electrical variables and basic features of the historical magnetic variables; performing feature crossover on the basic features of the historical electrical variables and the basic features of the historical magnetic variables to obtain crossover features of historical electromagnetic variables; performing Fourier transform and wavelet transform on the basic features of the historical electrical variables respectively to obtain historical electrical periodic features and historical electrical non-periodic features; performing Fourier transform and wavelet transform on the basic features of the historical magnetic variables respectively to obtain historical magnetic periodic features and historical magnetic non-periodic features.

[0009] Furthermore, the model structure of the power communication network fault prediction model includes an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer; the training and verification of the power communication network fault prediction model based on the historical electromagnetic comprehensive data set includes: determining a loss function and an optimizer, dividing the historical electromagnetic comprehensive data set into a training set and a test set; the input layer receives the training set and inputs the convolution layer through a multi-channel; the convolution layer extracts features and outputs a historical electromagnetic feature map; the pooling layer performs spatial dimensionality reduction on the historical electromagnetic feature map and outputs the historical electromagnetic feature map after dimensionality reduction; the fully connected layer applies the ReLU activation function to expand the historical electromagnetic feature map after dimensionality reduction to obtain the final historical electromagnetic feature map; the output layer applies the Sigmoid activation function to convert the final historical electromagnetic feature map into the probability of a power communication network failure, specifically extracting historical electromagnetic features from the final historical electromagnetic feature map, the historical electromagnetic features include historical electromagnetic amplitude and historical electromagnetic frequency, and processing the historical electromagnetic amplitude and historical electromagnetic frequency based on the Sigmoid activation function to obtain the probability of a power communication network failure corresponding to the final historical electromagnetic feature map, and the calculation formula is as follows:

[0010] ;

[0011] In the formula, is the probability of failure of the power communication network corresponding to the final historical electromagnetic characteristic map, is the historical electromagnetic amplitude, is the historical electromagnetic frequency, is a natural constant; obtaining an evaluation index of a fault prediction model for a power communication network, and if the evaluation index of the fault prediction model for a power communication network meets the standard, stopping the training; if the evaluation index of the fault prediction model for a power communication network meets the standard, continuing the iterative training; verifying the fault prediction model in the training process of the fault prediction model for a power communication network using the training set based on the test set, obtaining the test result, and tuning the model structure and hyperparameters of the fault prediction model for a power communication network based on the test result, wherein the hyperparameters include a learning rate, a number of iterations, and a regularization parameter, wherein the process of tuning the model structure and hyperparameters of the fault prediction model for a power communication network based on the test result is as follows: comparing the test result with the tuning scheme corresponding to each test result stored in a database, obtaining the tuning scheme corresponding to the successfully compared test result, and tuning the model structure and hyperparameters of the fault prediction model for a power communication network based on the tuning scheme.

[0012] Furthermore, the real-time electromagnetic comprehensive data set includes real-time basic features of electric variables, real-time basic features of magnetic variables, real-time cross features of electromagnetic variables, real-time electric periodic features, real-time electric non-periodic features, real-time magnetic periodic features and real-time magnetic non-periodic features; if the real-time fault probability of the electric power communication network is greater than or equal to the set fault threshold probability, the fault type of the electric power communication network is obtained based on the analysis of the real-time electromagnetic comprehensive data set, including: extracting and parsing the real-time electric periodic features and the real-time magnetic periodic features, and processing the real-time electric periodic features and the real-time magnetic periodic features with the fault electric periodic features and the fault magnetic periodic features corresponding to each fault type to obtain the electromagnetic periodic compliance coefficient; extracting and parsing the real-time electric non-periodic features and the real-time magnetic non-periodic features, and processing the real-time electric non-periodic features and the real-time magnetic non-periodic features with the fault electric non-periodic features and the fault magnetic non-periodic features corresponding to each fault type to obtain the electromagnetic non-periodic compliance coefficient; obtaining the fault compliance index based on the electromagnetic periodic compliance coefficient and the electromagnetic non-periodic compliance coefficient; if the fault compliance index corresponding to a certain fault type is less than or equal to the set fault compliance threshold, the fault type corresponding to the fault compliance index is the electric power communication network fault type.

[0013] Furthermore, the calculation formula of the fault compliance threshold is as follows:

[0014] ;

[0015] In the formula, For the fault compliance threshold, is the electromagnetic periodicity compliance coefficient, is the weight factor of the electromagnetic periodicity compliance coefficient, is the electromagnetic non-periodic coincidence coefficient, is the weighting factor of the electromagnetic non-periodic compliance coefficient.

[0016] Furthermore, the real-time electrical periodic characteristics include voltage fundamental frequency, current periodic load change and voltage-current phase relationship, and the real-time magnetic periodic characteristics include magnetic flux fundamental frequency, magnetic field strength periodic change and magnetic flux-magnetic field strength phase relationship; the fault electrical periodic characteristics include fault voltage fundamental frequency, fault current periodic load change and fault voltage-fault current phase relationship, and the fault magnetic periodic characteristics include fault magnetic flux fundamental frequency, fault magnetic field strength periodic change and fault magnetic flux-fault magnetic field strength phase relationship.

[0017] Furthermore, the real-time electrical non-periodic characteristics include voltage instantaneous impact, current transient disturbance and voltage frequency deviation, and the real-time magnetic non-periodic characteristics include magnetic flux instantaneous impact, magnetic field strength transient disturbance and magnetic flux frequency deviation; the fault electrical non-periodic characteristics include fault voltage instantaneous impact, fault current transient disturbance and fault voltage frequency deviation, and the fault magnetic non-periodic characteristics include fault magnetic flux instantaneous impact, fault magnetic field strength transient disturbance and fault magnetic flux frequency deviation.

[0018] A power communication network fault prediction system based on deep learning, used for the above-mentioned power communication network fault prediction method based on deep learning, comprising a historical electromagnetic data acquisition subsystem, a model training and verification subsystem, a fault probability prediction subsystem and a fault type determination subsystem, wherein: the historical electromagnetic data acquisition subsystem is used to integrate and process the acquired historical electrical variables and historical magnetic variables to obtain a historical electromagnetic comprehensive data set; the model training and verification subsystem is used to train and verify the power communication network fault prediction model based on the historical electromagnetic comprehensive data set, and the output of the power communication network fault prediction model is the probability of a power communication network failure; the fault probability prediction subsystem is used to predict the real-time electromagnetic comprehensive data set composed of real-time electrical variables and real-time magnetic variables based on the trained and verified power communication network fault prediction model, and obtain the real-time fault probability of the power communication network; the fault type determination subsystem is used to obtain the power communication network fault type based on the real-time electromagnetic comprehensive data set analysis if the real-time fault probability of the power communication network is greater than or equal to the set fault threshold probability.

[0019] A method and device for predicting faults in a power communication network based on deep learning, comprising a processor and a memory, wherein computer program instructions are stored in the memory, and when the computer program instructions are executed by the processor, the processor executes the method for predicting faults in a power communication network based on deep learning as described above, and further comprising: a historical electromagnetic data acquisition module, a model training and verification module, a fault probability prediction module, and a fault type determination module, wherein: the historical electromagnetic data acquisition module is used to integrate and process the acquired historical electrical variables and historical magnetic variables to obtain a historical electromagnetic comprehensive data set; the model training and verification module is used to train and verify a power communication network fault prediction model based on the historical electromagnetic comprehensive data set, and the output of the power communication network fault prediction model is the probability of a fault in the power communication network; the fault probability prediction module is used to predict a real-time electromagnetic comprehensive data set composed of real-time electrical variables and real-time magnetic variables based on the power communication network fault prediction model that has been trained and verified, and obtain the real-time fault probability of the power communication network; the fault type determination module is used to obtain the fault type of the power communication network based on the analysis of the real-time electromagnetic comprehensive data set if the real-time fault probability of the power communication network is greater than or equal to the set fault threshold probability.

[0020] The present invention has the following beneficial effects:

[0021] (1) This deep learning-based power communication network fault prediction method can capture the comprehensive impact of the power communication network status from multiple dimensions by simultaneously analyzing electrical and magnetic variables. Multi-dimensional analysis makes fault detection more comprehensive and can identify faults caused by specific electrical variables or magnetic variables, as well as the complex fault modes that may occur when the two interact with each other.

[0022] (2) This deep learning-based power communication network fault prediction method, combined with electrical and magnetic variables, can not only increase the data support for fault diagnosis, but also improve the accuracy and sensitivity of fault prediction by analyzing the periodic and non-periodic characteristics of these variables. The periodic characteristics help identify long-term or persistent anomalies, while the non-periodic characteristics focus on capturing instantaneous or transient fault events.

[0023] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of the power communication network fault prediction method based on deep learning of the present invention.

[0025] Figure 2 This is a flowchart of the power communication network fault prediction system based on deep learning of the present invention.

[0026] Figure 3 This is a structural connection diagram of the power communication network fault prediction device based on deep learning of the present invention. DETAILED DESCRIPTION

[0027] The embodiments of the present application use a power communication network fault prediction method, system and device based on deep learning. Through the combination of electrical variables and magnetic variables, the power communication network fault prediction method can provide more comprehensive, efficient and accurate fault diagnosis and prediction capabilities.

[0028] The overall idea of ​​the problem in the embodiment of this application is as follows:

[0029] First, the historically collected electrical and magnetic variable data are integrated to form a historical electromagnetic comprehensive dataset. This dataset will include the electromagnetic performance of the power communication network under different conditions, providing rich training samples for the deep learning model.

[0030] Using the integrated historical electromagnetic comprehensive data set, a deep learning model is trained to predict power communication network failures. The patterns and associations in the data are learned so that the probability of failure can be accurately predicted when real-time data is input in the future. During the model development stage, it also needs to be verified to ensure the accuracy and stability of its predictions.

[0031] The trained model is applied to the real-time collected electrical and magnetic variable data to form a real-time electromagnetic integrated data set, and the failure probability of the power communication network is predicted based on these real-time data. If the predicted failure probability reaches or exceeds the preset threshold, the data will also be analyzed to determine the possible fault type.

[0032] See also Figure 1 The embodiment of the present invention provides a technical solution: a method for predicting power communication network faults based on deep learning, comprising the following steps: integrating and processing the acquired historical electrical variables and historical magnetic variables to obtain a historical electromagnetic comprehensive data set; training and verifying a power communication network fault prediction model based on the historical electromagnetic comprehensive data set, wherein the output of the power communication network fault prediction model is the probability of a power communication network failure; predicting a real-time electromagnetic comprehensive data set composed of real-time electrical variables and real-time magnetic variables based on the trained and verified power communication network fault prediction model to obtain a real-time fault probability of the power communication network; if the real-time fault probability of the power communication network is greater than or equal to a set fault threshold probability, obtaining the power communication network fault type based on analysis of the real-time electromagnetic comprehensive data set.

[0033] By combining power and magnetic data, the power communication network fault prediction model can more comprehensively understand the status of the power communication network, thereby improving the accuracy of fault prediction. By predicting possible fault types, the operation and maintenance team can perform targeted maintenance or adjustments in advance to avoid faults and extend the service life of the equipment.

[0034] Specifically, the historical electromagnetic comprehensive data set includes basic characteristics of historical electric variables, basic characteristics of historical magnetic variables, cross characteristics of historical electromagnetic variables, historical electric periodic characteristics, historical electric non-periodic characteristics, historical magnetic periodic characteristics and historical magnetic non-periodic characteristics.

[0035] The historical electrical variables and historical magnetic variables are processed and integrated to obtain a historical electromagnetic comprehensive data set, including: unifying the timestamps and data formats of the historical electrical variables and historical magnetic variables, ensuring that all data are aligned in time series and have a consistent format, and synchronizing the time of all observation points in the data set, such as converting the recording time of all data points into a unified format (such as UTC timestamp). At the same time, ensuring that all data use the same digital representation, such as floating point numbers or integers. The unified timestamp and data format can reduce errors in the processing process and improve the efficiency and accuracy of data processing.

[0036] The basic characteristics of historical electric variables and historical magnetic variables are processed respectively to obtain the basic characteristics of historical electric variables and historical magnetic variables. The basic characteristics of historical electric variables include the calculated average value of voltage and current, maximum value of voltage and current, minimum value of voltage and current, standard deviation of voltage and current, etc. The basic characteristics of historical magnetic variables include the average value of magnetic flux, maximum value of magnetic field intensity, etc. These basic characteristics directly describe the basic statistical properties of each variable, provide a simple but effective data overview for the fault prediction model of the power communication network, and help to have a preliminary understanding of the data distribution and variability.

[0037] New features are created to reveal the interactions between different variables and enhance the predictive ability of the power communication network fault prediction model. The basic features of historical electric variables and the basic features of historical magnetic variables are cross-featured to obtain the cross-features of historical electromagnetic variables. The cross-features of historical electromagnetic variables include the product of current intensity and magnetic field intensity, etc. The cross-features can help the model capture the nonlinear relationship between electric and magnetic variables, which may not be obvious in simple single variable analysis. The basic features of historical electric variables are subjected to Fourier transform and wavelet transform respectively to obtain historical electric periodic features and historical electric non-periodic features. The basic features of historical magnetic variables are subjected to Fourier transform and wavelet transform respectively to obtain historical magnetic periodic features and historical magnetic non-periodic features.

[0038] In this implementation, through these steps, the power communication network fault prediction model can not only analyze and understand the data more comprehensively, but also predict and diagnose power communication network problems more accurately in practical applications. The comprehensive use of features provides a powerful, multi-dimensional data foundation for the deep learning model, greatly enhancing the accuracy and reliability of the prediction.

[0039] Specifically, the model structure of the power communication network fault prediction model includes an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer; the training and verification of the power communication network fault prediction model based on the historical electromagnetic comprehensive data set includes: determining a loss function and an optimizer, dividing the historical electromagnetic comprehensive data set into a training set and a test set; the input layer receives the training set and inputs the convolution layer through a multi-channel input, and the function of the input layer is to receive input data, that is, the historical electromagnetic comprehensive data set. In this model, the input layer is specially configured as a multi-channel input, which means that multiple types of electrical and magnetic variable data can be processed simultaneously. The multi-channel input allows the model to observe multiple data dimensions at the same time, which helps the model capture the intrinsic connection between different data sources.

[0040] The convolution layer extracts features from the input data by applying multiple filters (convolution kernels). These features are expressed as "feature maps" that reflect important information of the original input data and output historical electromagnetic feature maps. By extracting local features, the convolution layer can capture complex patterns and relationships in the electromagnetic data. The pooling layer performs spatial dimensionality reduction on the historical electromagnetic feature maps, usually using maximum pooling or average pooling methods to reduce the spatial size of the data, and outputs the historical electromagnetic feature maps after dimensionality reduction. The dimensionality reduction operation reduces the number of parameters and computational complexity while retaining important features, which helps prevent model overfitting and improve the generalization ability of the model.

[0041] The fully connected layer applies the ReLU activation function to expand the historical electromagnetic feature map after dimensionality reduction to obtain the final historical electromagnetic feature map. The fully connected layer expands the pooled feature map into a one-dimensional vector, and then further processes these features through one or more fully connected network layers. Here, the ReLU (linear rectifier unit) activation function is used to increase the nonlinear ability of the network so that the model can learn more complex functions; the output layer applies the Sigmoid activation function to the final historical electromagnetic feature map to convert it into the probability of a failure in the power communication network. The Sigmoid function is suitable for binary classification tasks and outputs a probability value between 0 and 1. Through the Sigmoid function, the model output can be directly interpreted as the probability of a failure, which makes the result intuitive and easy to compare with the threshold to make a decision. Specifically, the historical electromagnetic features are extracted from the final historical electromagnetic feature map. The historical electromagnetic features include historical electromagnetic amplitude and historical electromagnetic frequency. The historical electromagnetic amplitude and historical electromagnetic frequency are processed based on the Sigmoid activation function to obtain the probability of a failure in the power communication network corresponding to the final historical electromagnetic feature map. The calculation formula is as follows:

[0042] ;

[0043] In the formula, is the probability of failure of the power communication network corresponding to the final historical electromagnetic characteristic map, is the historical electromagnetic amplitude, is the historical electromagnetic frequency, is a natural constant; obtaining an evaluation index of the fault prediction model of the electric power communication network, if the evaluation index of the fault prediction model of the electric power communication network meets the standard, then stop the training; if the evaluation index of the fault prediction model of the electric power communication network does not meet the standard, then continue the iterative training; based on the test set, verify the fault prediction model in the training process of the fault prediction model of the electric power communication network using the training set, obtain the test result, and tune the structure and hyperparameters of the fault prediction model of the electric power communication network based on the test result, the hyperparameters include the learning rate, the number of iterations and the regularization parameter, wherein the process of tuning the model structure and hyperparameters of the fault prediction model of the electric power communication network based on the test result is: comparing the test result with the tuning scheme corresponding to each test result stored in the database, obtaining the tuning scheme corresponding to the successfully compared test result, and tuning the model structure and hyperparameters of the fault prediction model of the electric power communication network based on the tuning scheme (the tuning scheme specifically includes changing the number of convolutional layers and pooling layers, and modifying the learning rate, the number of iterations and the regularization parameter).

[0044] In this implementation, the power communication network fault prediction model is first trained on the training set, using a loss function (such as cross entropy) and an optimizer (such as Adam) to adjust the model weights. After that, the model is validated on the test set to evaluate its generalization ability. Separating the training set and the test set can verify the model's prediction ability for unseen data and ensure the reliability and practicality of the model. By iterative training and adjusting model parameters, the model performance can be continuously optimized.

[0045] Specifically, the real-time electromagnetic comprehensive data set includes real-time electric variable basic features, real-time magnetic variable basic features, real-time electromagnetic variable cross features, real-time electric periodic features, real-time electric non-periodic features, real-time magnetic periodic features and real-time magnetic non-periodic features; if the real-time fault probability of the electric power communication network is greater than or equal to the set fault threshold probability, the fault type of the electric power communication network is obtained based on the real-time electromagnetic comprehensive data set analysis, including: extracting and parsing real-time electric periodic features and real-time magnetic periodic features such as mean, variance, maximum value and minimum value, and corresponding real-time electric periodic features and real-time magnetic periodic features to each fault type The fault electrical periodic characteristics and fault magnetic periodic characteristics are processed to obtain the electromagnetic periodic compliance coefficient; the real-time electrical non-periodic characteristics and real-time magnetic non-periodic characteristics are extracted and analyzed, and the real-time electrical non-periodic characteristics and real-time magnetic non-periodic characteristics are processed with the fault electrical non-periodic characteristics and fault magnetic non-periodic characteristics corresponding to each fault type to obtain the electromagnetic non-periodic compliance coefficient; the fault compliance index is obtained based on the electromagnetic periodic compliance coefficient and the electromagnetic non-periodic compliance coefficient; if the fault compliance index corresponding to a certain fault type is less than or equal to the set fault compliance threshold, the fault type corresponding to the fault compliance index is the power communication network fault type.

[0046] Extract the basic statistical features of real-time electrical and magnetic variables, such as mean, variance, maximum and minimum values, which provide direct insights into the current status of the power communication network and help the model quickly capture the baseline state of the power communication network. Combine the features of electrical and magnetic variables to form cross-features, such as the product of current intensity and magnetic field intensity, in this way, the interaction between electrical and magnetic features can be explored.

[0047] Extracting periodic and non-periodic features from real-time data using methods such as Fourier transform and wavelet transform helps identify regularities and abnormal behaviors in the operation of the power communication network. Compare the real-time electromagnetic periodic features with the periodic features of known fault types and calculate the compliance coefficient. Compare the real-time electromagnetic non-periodic features with the non-periodic features of the fault type and calculate the compliance coefficient in the same way.

[0048] The fault compliance index is calculated based on the electromagnetic periodicity and non-periodicity compliance coefficients and compared with the preset fault compliance threshold. If the real-time calculated power communication network fault probability exceeds the preset threshold, the fault analysis process will be triggered. The fault compliance index provides a quantitative indicator to evaluate the relevance of each fault type. If the fault compliance index is lower than the threshold, the corresponding fault type is confirmed to be a problem of the current power communication network. For the set fault compliance threshold, the fault compliance threshold can be estimated by analyzing historical data.

[0049] The calculation formula of the fault compliance threshold is as follows:

[0050] ;

[0051] In the formula, For the fault compliance threshold, is the electromagnetic periodicity compliance coefficient, is the weight factor of the electromagnetic periodicity compliance coefficient, is the electromagnetic non-periodic coincidence coefficient, is the weighting factor of the electromagnetic non-periodic compliance coefficient.

[0052] In this implementation, in the operation of the power communication network, some faults may be closely related to periodic changes (such as load fluctuations), while other faults may be related to non-periodic events (such as equipment failures). Combining these two types of features can more comprehensively cover various fault situations. Periodic and non-periodic features represent different aspects of the power communication network status on a time scale, and can effectively identify and respond to long-term and short-term abnormal changes.

[0053] By adjusting the weight factors, the accuracy of fault detection can be optimized according to different power communication network environments and fault characteristics, making the model more adaptable. The method of combining periodic and non-periodic characteristics enables the model to take into account both the long-term trends and short-term fluctuations of power communication network data, thereby improving the comprehensiveness and accuracy of fault prediction.

[0054] For the weight factor and , can be obtained through data analysis and optimization, and statistical analysis can be performed, such as calculating the correlation coefficients between periodic and non-periodic features and fault events to preliminarily evaluate their respective importance, and using machine learning techniques such as gradient descent or evolutionary algorithms to adjust weights on the validation set to maximize a certain performance indicator (such as accuracy, recall).

[0055] Specifically, the real-time electrical periodic characteristics include voltage basic frequency, current periodic load changes and voltage-current phase relationship, and the real-time magnetic periodic characteristics include magnetic flux basic frequency, magnetic field strength periodic changes and magnetic flux-magnetic field strength phase relationship; the fault electrical periodic characteristics include fault voltage basic frequency, fault current periodic load changes and fault voltage-fault current phase relationship, and the fault magnetic periodic characteristics include fault magnetic flux basic frequency, fault magnetic field strength periodic changes and fault magnetic flux-fault magnetic field strength phase relationship.

[0056] Voltage fundamental frequency and fault voltage fundamental frequency use Fourier transform to extract the fundamental frequency from the voltage signal. This analysis is performed on the data records under normal and fault conditions to obtain the fundamental frequencies under real-time and fault conditions, respectively. Based on the analysis of historical fault data, the typical voltage fundamental frequency at the time of fault is determined. Current periodic load change and fault current periodic load change are analyzed by analyzing the time series data of the current signal, using Fourier transform or wavelet transform to identify periodic load fluctuations, including calculating the amplitude and frequency of periodic fluctuations. Based on historical data, the periodic load change characteristics of the current under fault conditions are statistically analyzed.

[0057] The voltage-current phase relationship and the fault voltage-current phase relationship are determined by cross-correlation analysis of the voltage and current waveforms or by directly measuring the phase angle between the two. The synchronously sampled voltage and current signals can be used for calculation, historical fault data can be analyzed, and the voltage and current phase difference at the time of the fault can be calibrated.

[0058] The fundamental frequency of magnetic flux and the fundamental frequency of fault magnetic flux are similar to the determination of the fundamental frequency of voltage. The fundamental frequency of magnetic flux is extracted from the output signal of the magnetic field sensor using Fourier transform, and the fundamental frequency of magnetic flux is extracted from the record of known fault events and used as a reference for fault mode. The periodic change of magnetic field intensity and the periodic change of fault magnetic field intensity are defined by monitoring the change of magnetic field intensity over time and applying Fourier or wavelet transform to analyze its periodic characteristics and analyzing the change pattern of magnetic field intensity in fault data to define fault characteristics.

[0059] The magnetic flux-magnetic field strength phase relationship and the fault magnetic flux-magnetic field strength phase relationship are determined based on the fault record data by measuring the waveforms of the magnetic flux and magnetic field strength and calculating the phase difference between the two to determine the phase relationship between the magnetic flux and magnetic field strength under the fault state.

[0060] ;

[0061] In the formula, is the number of each feature in the real-time electrical periodic feature or the real-time magnetic periodic feature or the fault electrical periodic feature or the fault magnetic periodic feature, ,when hour, is the voltage fundamental frequency, is the fault voltage fundamental frequency, is the voltage base frequency weight factor, is the fundamental frequency of magnetic flux, is the fundamental frequency of the fault magnetic flux, is the flux base frequency weighting factor, when hour, For the periodic load change of current, The fault current is periodic load variation. is the current periodic load change weight factor, is the periodic change of magnetic field intensity, The fault magnetic field strength changes periodically. is the weight factor of periodic change of magnetic field intensity, when hour, is the voltage-current phase relationship, is the fault voltage-fault current phase relationship, is the voltage-current phase relationship weight factor, is the phase relationship between magnetic flux and magnetic field intensity, is the phase relationship between fault magnetic flux and fault magnetic field intensity, is the weight factor of the phase relationship between magnetic flux and magnetic field intensity, A is the weight factor of the real-time electrical periodicity characteristic, B is the weight factor of the real-time magnetic periodicity characteristic, , , .

[0062] In this implementation scheme, the basic frequency, periodic changes and phase relationship of electricity and magnetism are key parameters commonly used in power communication network analysis. These parameters directly reflect the physical state and operating efficiency of the power communication network. By comparing the differences between real-time characteristics and fault characteristics, signals that deviate from normal operating conditions can be intuitively identified.

[0063] The basic periodic features of electricity and magnetism and their corresponding features in the fault condition are selected, including fundamental frequency, periodic load changes and phase relationships, each with its own specific weight factor, which reflects the importance of each feature in fault detection. By calculating the difference between the real-time features and the corresponding fault features and taking a weighted sum of these differences, the deviation between the real-time data and the known fault condition can be quantified. The weighted feature differences are used to calculate the overall periodic compliance coefficient, which takes into account the electrical and magnetic features to form a comprehensive fault detection index.

[0064] By considering multiple periodic characteristics of electricity and magnetism and their interrelationships, the dynamic changes of the power communication network can be fully captured, the accuracy of fault prediction is increased, and the early warning capability of fault detection is improved by sensitively responding to subtle differences between real-time data and fault data.

[0065] The weight factors can be optimized during the training phase using machine learning methods, such as minimizing the prediction error through techniques such as gradient descent.

[0066] Specifically, the real-time electrical non-periodic features include instantaneous voltage surges, transient current disturbances and voltage frequency deviations, which can reveal sudden and short-term abnormal phenomena in the power communication network, such as instantaneous overload, circuit break or equipment failure. The real-time magnetic non-periodic features include instantaneous magnetic flux surges, transient disturbances of magnetic field strength and magnetic flux frequency deviations. Similarly, these magnetic-related non-periodic features indicate mutations in the magnetic field, which are caused by power communication network faults; the fault electrical non-periodic features include fault voltage transient surges, fault current transient disturbances and fault voltage frequency deviations, and the fault magnetic non-periodic features include fault magnetic flux transient surges, fault magnetic field strength transient disturbances and fault magnetic flux frequency deviations.

[0067] Voltage transient impact and fault voltage transient impact use voltage sensors to monitor voltage waveforms in real time, using high-speed sampling technology to capture instantaneous spikes or impacts. Fault voltage transient impacts can be analyzed from historical fault data, usually the voltage spike value observed during the fault. Current transient disturbances and fault current transient disturbances use current sensors to monitor current waveforms, especially rapid changes in current in a short period of time. Fault characteristics are based on the typical transient response of current in similar faults that occurred in the past. Voltage frequency deviation and fault voltage frequency deviation analyze the frequency of the voltage signal in real time and compare it with a standard frequency (such as 50 Hz or 60 Hz) to obtain the deviation. Fault voltage frequency deviation is a statistical data of frequency deviation obtained from historical fault events.

[0068] Magnetic flux transient impact and fault magnetic flux transient impact monitor magnetic flux through magnetic field sensors and record transient impact events. Fault magnetic flux impact is based on abnormal peak values ​​of magnetic field in historical data. Magnetic field strength transient disturbance and fault magnetic field strength transient disturbance monitor the changes of magnetic field strength in real time, especially the rapid changes. Historical fault data analysis provides typical transient changes of magnetic field during faults. Magnetic flux frequency deviation and fault magnetic flux frequency deviation analyze the frequency changes of magnetic field signals and compare them with the frequency under standard operating conditions. Fault magnetic flux frequency deviation is obtained by analyzing fault records.

[0069] ;

[0070] In the formula, is the number of each feature in the real-time electrical periodic feature or the real-time magnetic periodic feature or the fault electrical periodic feature or the fault magnetic periodic feature, ,when hour, is the voltage instantaneous impact, is the instantaneous impact of fault voltage, is the voltage transient impact weight factor, is the instantaneous impact of magnetic flux, is the instantaneous impact of fault magnetic flux, is the instantaneous impact weight factor of magnetic flux, when hour, is the current transient disturbance, is the fault current transient disturbance, is the current transient disturbance weight factor, is the transient disturbance of magnetic field intensity, is the transient disturbance of fault magnetic field strength, is the transient disturbance weight factor of magnetic field intensity, when hour, is the voltage-frequency deviation, is the fault voltage frequency deviation, is the voltage-frequency deviation weight factor, is the fault flux frequency deviation, is the fault flux frequency deviation, is the flux frequency deviation weight factor, is the real-time electrical non-periodic characteristic weight factor, is the real-time magnetic non-periodic characteristic weight factor, , , .

[0071] In this implementation, some fault types may be mainly manifested through non-periodic changes, such as instantaneous shocks caused by equipment short circuits, and periodic characteristics may not fully cover all types of faults, especially those transient or discontinuous faults. Electrical parameters and magnetic parameters complement each other in function and together provide a comprehensive view of the state of the power communication network. By integrating these features and weighting factors, the deviation between real-time data and fault modes can be more accurately evaluated, thereby improving the accuracy and response speed of fault detection.

[0072] The difference between each real-time feature and its corresponding fault feature is calculated and weighted to obtain the total non-periodic compliance coefficient. The non-periodic features enable the model to respond quickly to emergencies in the power communication network and improve the timeliness of fault detection. By analyzing these instantaneous and transient features, faults that may be overlooked in periodic analysis can be more sensitively identified.

[0073] A power communication network fault prediction system based on deep learning, used for the above-mentioned power communication network fault prediction method based on deep learning, such as Figure 2 As shown, it includes a historical electromagnetic data acquisition subsystem, a model training and verification subsystem, a fault probability prediction subsystem and a fault type determination subsystem, wherein: the historical electromagnetic data acquisition subsystem is used to integrate and process the acquired historical electrical variables and historical magnetic variables to obtain a historical electromagnetic comprehensive data set; the model training and verification subsystem is used to train and verify the power communication network fault prediction model based on the historical electromagnetic comprehensive data set, and the output of the power communication network fault prediction model is the probability of a power communication network failure; the fault probability prediction subsystem is used to predict the real-time electromagnetic comprehensive data set composed of real-time electrical variables and real-time magnetic variables based on the power communication network fault prediction model that has been trained and verified to obtain the real-time fault probability of the power communication network; the fault type determination subsystem is used to obtain the power communication network fault type based on the real-time electromagnetic comprehensive data set analysis if the real-time fault probability of the power communication network is greater than or equal to the set fault threshold probability.

[0074] A method and device for predicting faults in a power communication network based on deep learning, comprising a processor and a memory, wherein computer program instructions are stored in the memory, and when the computer program instructions are executed by the processor, the processor executes the method for predicting faults in a power communication network based on deep learning, such as Figure 3 As shown, it also includes: a historical electromagnetic data acquisition module, a model training and verification module, a fault probability prediction module and a fault type determination module, wherein: the historical electromagnetic data acquisition module is used to integrate and process the acquired historical electrical variables and historical magnetic variables to obtain a historical electromagnetic comprehensive data set; the model training and verification module is used to train and verify the power communication network fault prediction model based on the historical electromagnetic comprehensive data set, and the output of the power communication network fault prediction model is the probability of a failure in the power communication network; the fault probability prediction module is used to predict the real-time electromagnetic comprehensive data set composed of real-time electrical variables and real-time magnetic variables based on the power communication network fault prediction model completed by training and verification, and obtain the real-time fault probability of the power communication network; the fault type determination module is used to obtain the power communication network fault type based on the real-time electromagnetic comprehensive data set analysis if the real-time fault probability of the power communication network is greater than or equal to the set fault threshold probability.

[0075] A computer-readable storage medium is used to store a program, which, when executed by a processor, implements the power communication network fault prediction method based on deep learning as described above.

[0076] In summary, this application has at least the following effects:

[0077] By analyzing electrical and magnetic variables simultaneously, the combined impact of the state of the power communication network can be captured from multiple dimensions. Multi-dimensional analysis makes fault detection more comprehensive and can identify faults caused by specific electrical or magnetic variables, as well as the complex fault modes that may occur when the two interact. Combining electrical and magnetic variables can not only increase the data support for fault diagnosis, but also improve the accuracy and sensitivity of fault prediction by analyzing the periodic and non-periodic characteristics of these variables. Periodic characteristics help identify long-term or persistent anomalies, while non-periodic characteristics focus on capturing instantaneous or transient fault events.

[0078] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0079] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0080] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0081] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0082] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0083] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting power communication network faults based on deep learning, characterized in that: The following steps are involved: The acquired historical electrical variables and historical magnetic variables are integrated and processed to obtain a historical electromagnetic comprehensive data set; Training and verifying a power communication network fault prediction model based on a historical electromagnetic comprehensive data set, wherein the output of the power communication network fault prediction model is a probability of a power communication network failure; Based on the trained and verified power communication network fault prediction model, the real-time electromagnetic comprehensive data set composed of real-time electrical variables and real-time magnetic variables is predicted to obtain the real-time fault probability of the power communication network; If the real-time fault probability of the power communication network is greater than or equal to the set fault threshold probability, the fault type of the power communication network is obtained based on the real-time electromagnetic comprehensive data set analysis; The historical electromagnetic comprehensive data set includes historical electric variable basic characteristics, historical magnetic variable basic characteristics, historical electromagnetic variable cross characteristics, historical electric periodic characteristics, historical electric non-periodic characteristics, historical magnetic periodic characteristics and historical magnetic non-periodic characteristics; The historical electrical variables and historical magnetic variables are processed and integrated to obtain a historical electromagnetic comprehensive data set, including: The acquired historical electrical variables and historical magnetic variables are time stamped and data formatted in a unified manner; Performing basic feature processing on historical electric variables and historical magnetic variables respectively to obtain basic features of historical electric variables and basic features of historical magnetic variables; Perform feature crossover on historical electric variable basic features and historical magnetic variable basic features to obtain historical electromagnetic variable crossover features; The characteristic crossover is the product of the current intensity and the magnetic field intensity; The basic characteristics of historical electrical variables are respectively subjected to Fourier transform and wavelet transform to obtain historical electrical periodic characteristics and historical electrical non-periodic characteristics; The basic characteristics of historical magnetic variables are subjected to Fourier transform and wavelet transform respectively to obtain historical magnetic periodic characteristics and historical magnetic non-periodic characteristics.

2. The method for predicting power communication network faults based on deep learning according to claim 1, characterized in that: The model structure of the power communication network fault prediction model includes an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer; The training and verification of the power communication network fault prediction model based on the historical electromagnetic comprehensive data set includes: Determine the loss function and optimizer, and divide the historical electromagnetic synthetic dataset into a training set and a test set; The input layer receives the training set and inputs it into the convolution layer through multiple channels; The convolutional layer extracts features and outputs historical electromagnetic feature maps; The pooling layer performs spatial dimensionality reduction on the historical electromagnetic characteristic map and outputs the historical electromagnetic characteristic map after dimensionality reduction; The fully connected layer uses the ReLU activation function to expand the historical electromagnetic characteristic map after dimension reduction to obtain the final historical electromagnetic characteristic map; The output layer applies the Sigmoid activation function to the final historical electromagnetic characteristic map and converts it into the probability of a power communication network failure. Specifically, the historical electromagnetic features are extracted from the final historical electromagnetic characteristic map. The historical electromagnetic features include the historical electromagnetic amplitude and the historical electromagnetic frequency. The historical electromagnetic amplitude and the historical electromagnetic frequency are processed based on the Sigmoid activation function to obtain the probability of a power communication network failure corresponding to the final historical electromagnetic characteristic map. The calculation formula is as follows: ； In the formula, is the probability of failure of the power communication network corresponding to the final historical electromagnetic characteristic map, is the historical electromagnetic amplitude, is the historical electromagnetic frequency, is a natural constant; Obtaining the fault prediction model evaluation index during the training process of the power communication network fault prediction model using the training set, if the evaluation index of the power communication network fault prediction model meets the standard, then stop the training, if the evaluation index of the power communication network fault prediction model does not meet the standard, then continue iterative training; The electric power communication network fault prediction model is verified based on the test set, the test results are obtained, and the model structure and hyperparameters of the electric power communication network fault prediction model are tuned based on the test results. The hyperparameters include learning rate, number of iterations and regularization parameters. The process of tuning the model structure and hyperparameters of the electric power communication network fault prediction model based on the test results is: comparing the test results with the tuning schemes corresponding to the test results stored in the database, obtaining the tuning schemes corresponding to the successfully compared test results, and tuning the model structure and hyperparameters of the electric power communication network fault prediction model based on the tuning schemes.

3. The method for predicting power communication network faults based on deep learning according to claim 1, characterized in that: The real-time electromagnetic comprehensive data set includes real-time electric variable basic features, real-time magnetic variable basic features, real-time electromagnetic variable cross features, real-time electric periodic features, real-time electric non-periodic features, real-time magnetic periodic features and real-time magnetic non-periodic features; If the real-time fault probability of the power communication network is greater than or equal to the set fault threshold probability, the fault type of the power communication network is obtained based on the real-time electromagnetic comprehensive data set analysis, including: Extract and analyze the real-time electrical periodicity characteristics and the real-time magnetic periodicity characteristics, and process the real-time electrical periodicity characteristics and the real-time magnetic periodicity characteristics with the fault electrical periodicity characteristics and the fault magnetic periodicity characteristics corresponding to each fault type to obtain the electromagnetic periodicity compliance coefficient; Extract and analyze the real-time electrical non-periodic features and the real-time magnetic non-periodic features, and process the real-time electrical non-periodic features and the real-time magnetic non-periodic features with the fault electrical non-periodic features and the fault magnetic non-periodic features corresponding to each fault type to obtain the electromagnetic non-periodic compliance coefficient; A fault compliance index is obtained based on the electromagnetic periodic compliance coefficient and the electromagnetic non-periodic compliance coefficient; If the fault compliance index corresponding to a certain fault type is less than or equal to the set fault compliance threshold, the fault type corresponding to the fault compliance index is the power communication network fault type.

4. A method for predicting faults in a power communication network based on deep learning according to claim 3, characterized in that: The calculation formula of the fault compliance threshold is as follows: ； In the formula, For the fault compliance threshold, is the electromagnetic periodicity compliance coefficient, is the weight factor of the electromagnetic periodicity compliance coefficient, is the electromagnetic non-periodic coincidence coefficient, is the weighting factor of the electromagnetic non-periodic compliance coefficient.

5. A method for predicting faults in a power communication network based on deep learning according to claim 4, characterized in that: The real-time electrical periodic characteristics include voltage fundamental frequency, current periodic load variation and voltage-current phase relationship, and the real-time magnetic periodic characteristics include magnetic flux fundamental frequency, magnetic field intensity periodic variation and magnetic flux-magnetic field intensity phase relationship; The fault electrical periodic characteristics include the fault voltage basic frequency, the fault current periodic load change and the fault voltage-fault current phase relationship, and the fault magnetic periodic characteristics include the fault magnetic flux basic frequency, the fault magnetic field intensity periodic change and the fault magnetic flux-fault magnetic field intensity phase relationship.

6. A method for predicting power communication network faults based on deep learning according to claim 4, characterized in that: The real-time electrical non-periodic characteristics include voltage transient impact, current transient disturbance and voltage frequency deviation, and the real-time magnetic non-periodic characteristics include magnetic flux transient impact, magnetic field strength transient disturbance and magnetic flux frequency deviation; The fault electrical non-periodic characteristics include fault voltage instantaneous impact, fault current transient disturbance and fault voltage frequency deviation, and the fault magnetic non-periodic characteristics include fault magnetic flux instantaneous impact, fault magnetic field strength transient disturbance and fault magnetic flux frequency deviation.

7. A power communication network fault prediction system based on deep learning, used in a power communication network fault prediction method based on deep learning according to any one of claims 1 to 6, characterized in that: It includes historical electromagnetic data acquisition subsystem, model training and verification subsystem, fault probability prediction subsystem and fault type determination subsystem, among which: The historical electromagnetic data acquisition subsystem is used to integrate and process the acquired historical electrical variables and historical magnetic variables to obtain a historical electromagnetic comprehensive data set; The model training and verification subsystem is used to train and verify the power communication network fault prediction model based on the historical electromagnetic comprehensive data set, and the output of the power communication network fault prediction model is the probability of failure of the power communication network; The fault probability prediction subsystem is used to predict the real-time electromagnetic comprehensive data set composed of real-time electrical variables and real-time magnetic variables based on the trained and verified power communication network fault prediction model to obtain the real-time fault probability of the power communication network; The fault type determination subsystem is used to obtain the fault type of the power communication network based on the real-time electromagnetic comprehensive data set analysis if the real-time fault probability of the power communication network is greater than or equal to the set fault threshold probability.

8. A power communication network fault prediction device based on deep learning, characterized in that: The method comprises a processor and a memory, wherein computer program instructions are stored in the memory, and when the computer program instructions are executed by the processor, the processor executes the power communication network fault prediction method based on deep learning according to any one of claims 1 to 6, and further comprises: a historical electromagnetic data acquisition module, a model training and verification module, a fault probability prediction module, and a fault type determination module, wherein: The historical electromagnetic data acquisition module is used to integrate the acquired historical electrical variables and historical magnetic variables to obtain a historical electromagnetic comprehensive data set; The model training and verification module is used to train and verify the power communication network fault prediction model based on the historical electromagnetic comprehensive data set, and the output of the power communication network fault prediction model is the probability of failure of the power communication network; The fault probability prediction module is used to predict the real-time electromagnetic comprehensive data set composed of real-time electrical variables and real-time magnetic variables based on the trained and verified power communication network fault prediction model to obtain the real-time fault probability of the power communication network; The fault type determination module is used to obtain the fault type of the power communication network based on the real-time electromagnetic comprehensive data set analysis if the real-time fault probability of the power communication network is greater than or equal to the set fault threshold probability.

Images