Power frequency waveform classification method and device, medium and equipment

Abstract

The invention discloses a power frequency waveform classification method and device, a medium and equipment, and belongs to the field of waveform classification, and the method comprises the steps: obtaining preprocessed power frequency waveform data of a power transmission line, calculating various waveform features, and inputting the features into a preset cascade classifier model. The model is composed of a plurality of XGBoost binary classifiers which are cascaded in sequence, and each classifier is responsible for executing a specific binary classification task. The output result of the previous classifier is used for triggering the starting of the next classifier, or the final classification result is directly output. The cascade structure not only improves the accuracy and efficiency of classification, but also reduces the complexity of the model, and effectively solves the problem that the fault positioning power frequency of the power transmission line cannot be accurately and efficiently classified in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of waveform classification, and in particular to a power frequency waveform classification method, device, medium and equipment. BACKGROUND

[0002] With the continuous expansion of the scale of the power system and the increase of the proportion of new energy grid connection, as the core infrastructure of the power grid, the transmission and distribution lines are affected by lightning, equipment aging, external damage and other factors, and fault tripping accidents occur frequently, which seriously threatens the power supply reliability and user safety of the power grid. The traditional fault processing relies on manual inspection, which not only requires a large amount of manpower and material resources, but also has the problems of long time-consuming and low accuracy in fault positioning, often leading to prolonged power outage and exacerbating economic losses and impact on people's livelihood.

[0003] In order to break through the limitations of manual inspection, the industry has gradually promoted the distributed fault positioning technology of transmission lines, which deploys fault positioning devices along the line, collects line power frequency current waveform data in real time and uploads it to the master station system, and analyzes the waveform data by the automatic fault diagnosis function of the master station to quickly determine the fault section, accurate position and fault type, greatly improving the operation and maintenance efficiency and fault disposal speed. In this process, power frequency waveform classification is the core prerequisite for automatic fault diagnosis: whether the fault is triggered, how to divide the fault section, and how to determine the fault type, all of which require accurate waveform classification results as the basis, therefore the accuracy, operation efficiency and resource consumption of power frequency waveform classification directly determine the practicality and reliability of the entire fault positioning system.

[0004] The current mainstream power frequency waveform classification method is mostly based on deep learning technologies such as long short-term memory artificial neural network (LSTM) and recurrent neural network (RNN). Although this type of technology can learn waveform features through data-driven mode, it has significant technical shortcomings: first, the training and running of deep learning models require professional deep learning frameworks, and the model complexity is high, requiring much more computer graphics processing unit (GPU) computing power than other modules of the automatic fault diagnosis system, which is likely to form a performance bottleneck due to excessive resource consumption in actual deployment, and cannot meet the real-time requirements of fault positioning; second, existing methods generally use a single model to perform multiple classification tasks, which need to distinguish between normal and abnormal waveforms, and the core waveform features for classification are not clearly defined, so the model needs to explore the differences between different waveforms. When the amount of training data is limited or the waveform scene is complex, not only is the classification difficulty increased, but also the generalization performance is poor, such as the training set is well adapted, but the test set is poorly adapted, which makes it difficult to cope with complex situations such as waveform interference and amplitude fluctuations in actual power grids.

[0005] These shortcomings result in the inability of existing technologies to accurately and efficiently classify power frequency for fault positioning of transmission lines. SUMMARY

[0006] The application provides a power frequency waveform classification method and device, a medium and equipment to solve the problem that the prior art cannot accurately and efficiently classify power frequency for fault positioning of a power transmission line.

[0007] In a first aspect, the application provides a power frequency waveform classification method, comprising: obtaining preprocessed power transmission line power frequency waveform data to be classified; based on the power transmission line power frequency waveform data, calculating various waveform features of the power transmission line power frequency waveform data; inputting the various waveform features into a preset cascade classifier model, so that the cascade classifier model performs a binary classification task according to a plurality of XGBoost binary classifiers which are sequentially cascaded internally, and the classification result of a previous XGBoost binary classifier is used to trigger a subsequent XGBoost binary classifier to start or directly output a power frequency waveform classification result; wherein the cascade classifier model is obtained by sequentially cascading a plurality of initial XGBoost binary classifiers which are respectively trained based on historical power frequency waveform data labeled with preset waveform types and the various waveform features; and the preset waveform types include a first preset waveform, a second preset waveform, an incoming waveform, a power-off waveform, a tripping waveform, a string-in waveform and a disturbance waveform.

[0008] The application can eliminate noise interference and format differences (such as uniform data volume) of original data by obtaining preprocessed power frequency waveform data of a power transmission line, provide consistent input for subsequent feature calculation, avoid classification deviation caused by irregular data, and lay the foundation for accurate classification. Secondly, based on data, the power frequency waveform data of the power transmission line is calculated for each waveform feature (such as average value, frequency domain energy, and wave peak number). These features are designed for the core differences of different waveform types (such as amplitude anomaly of the second preset waveform and three-section amplitude change of the tripping waveform). The model does not need to explore the features independently, can directly focus on the key distinguishing dimensions, reduces the invalid learning cost, improves the interpretability of the features, and provides support for improving the classification accuracy. Furthermore, the features are input into a model in which multiple XGBoost binary classifiers are sequentially cascaded. Each classifier only performs a specific binary classification task (such as distinguishing normal / second preset waveform and then subdividing steady-state / non-steady-state waveform). Compared with a traditional single multi-classification model, the task complexity of a single classifier is greatly reduced, making it easier for the model to learn the subtle differences between various waveforms. The logic of the previous stage classification result triggering the next stage classification (such as directly outputting the second preset waveform without subsequent classification) can reduce redundant calculation and improve classification efficiency. At the same time, the XGBoost model does not need to rely on a deep learning framework and has no special requirements for GPU computing power, which can significantly reduce resource consumption during training and operation and avoid performance bottlenecks. Finally, the cascaded classifier is trained based on historical data with preset waveform types and preset features, which can fully utilize the typical rules of historical waveforms, so that the model can still stably classify when facing complex waveforms such as tripping and stringing, and improve the generalization ability. In summary, through the technical logic of "standardized data-clear features-staged classification-lightweight architecture", the application can simultaneously improve the accuracy, efficiency, and generalization ability of power frequency waveform classification while reducing resource consumption, providing reliable support for automatic diagnosis of power transmission line fault positioning, and optimizing the efficiency of power grid fault handling. The application effectively solves the problem of accurately and efficiently classifying power frequency for power transmission line fault positioning in the prior art.

[0009] Further, the preprocessed power frequency waveform data to be classified is obtained, specifically: obtaining original power frequency waveform data to be classified collected by a distributed fault positioning device of a power transmission line; performing median filtering processing on the original power frequency waveform data to be classified to obtain filtered waveform data; if the data volume of the filtered waveform data is less than a preset data volume, data points are supplemented by interpolation; if the data volume of the filtered waveform data is greater than the preset data volume, data points are reduced by downsampling, so that the filtered waveform data meets the preset data standard, and the preprocessed power frequency waveform data to be classified of the power transmission line is obtained.

[0010] In the process of obtaining power frequency waveform data of a power transmission line to be classified, the application first collects original power frequency waveform data through a distributed fault location device of the power transmission line, and then performs median filtering processing on the original power frequency waveform data to remove noise and ensure the purity of the data. Then, according to the comparison result of the data amount of the filtered waveform data and the preset data amount, the data point amount is adjusted by interpolation or decimation, so that the data meets the preset standard. This process not only effectively removes the noise in the waveform data, but also ensures that the subsequent classifier can process data in a unified format through the standardization processing of the data amount, thereby improving the consistency and usability of the data. At the same time, the interpolation and decimation method optimizes the data amount under the premise of ensuring the data quality, reduces unnecessary computational burden, and improves the operation efficiency. Therefore, this preprocessing method provides high-quality and standardized data input for the subsequent power frequency waveform classification, thereby improving the accuracy and efficiency of the classification.

[0011] Further, based on the power frequency waveform data of the power transmission line, each waveform feature of the power frequency waveform data of the power transmission line is calculated, specifically: The each waveform feature includes mean, standard deviation, amplitude range, high frequency energy, medium frequency energy, low frequency energy, smoothness feature, peak number and stationarity feature; The mean, standard deviation and amplitude range are obtained by performing arithmetic operations on the amplitude of the power frequency waveform data of the power transmission line; The high frequency energy, medium frequency energy and low frequency energy are obtained by converting the power frequency waveform data of the power transmission line into frequency domain, dividing different frequency intervals, and accumulating the energy components of each frequency interval; The peak number is obtained based on the amplitude variation law of the power frequency waveform data of the power transmission line; The smoothness feature is calculated by the difference between the original power frequency waveform data to be classified and the power frequency waveform data of the power transmission line, and the stationarity feature is obtained by dividing the data window and analyzing the stability of the data in the window.

[0012] When calculating each waveform feature of the power frequency waveform data of the power transmission line, the application selects a plurality of features including mean, standard deviation, amplitude range, high frequency energy, medium frequency energy, low frequency energy, smoothness feature, peak number and stationarity feature. These features reflect the characteristics of the power frequency waveform from different angles and can effectively distinguish different types of power frequency waveform. The mean, standard deviation and amplitude range are obtained by arithmetic operation, which can understand the basic statistical characteristics of the waveform; the energy components of different frequency intervals are obtained by frequency domain conversion, which can analyze the frequency characteristics of the waveform; the peak number is obtained by statistics, which can capture the mutation of the waveform; the smoothness feature and the stationarity feature are calculated, which can evaluate the smoothness and stability of the waveform. The comprehensive use of these features enables the classifier to more accurately identify different types of power frequency waveform, thereby improving the accuracy and reliability of the classification.

[0013] Further, the various waveform features are input into a preset cascade classifier model, so that the cascade classifier model performs a binary classification task according to a plurality of XGBoost binary classifiers internally and sequentially cascaded, the classification result of a previous XGBoost binary classifier is used to trigger a subsequent XGBoost binary classifier to start or directly output a power frequency waveform classification result, specifically: The various waveform features are input into a preset cascade classifier model, so that the cascade classifier model first distinguishes a first preset waveform from a second preset waveform through a first XGBoost binary classifier, and if the classification result is the second preset waveform, the classification result of the second preset waveform is output. If the classification result is the first preset waveform, a second XGBoost binary classifier is used to distinguish a first preset steady-state waveform from a non-first preset steady-state waveform, and if the classification result is the first preset steady-state waveform, the classification result of the first preset steady-state waveform is output. If the classification result is a non-first preset steady-state waveform, a third XGBoost binary classifier is used to distinguish a two-segment waveform from a three-segment waveform. If it is a two-segment waveform, a fourth XGBoost binary classifier is used to distinguish a power-on waveform from a power-off waveform and output the classification result of the power-on waveform or the power-off waveform. If it is a three-segment waveform, a fifth XGBoost binary classifier is used to distinguish a trip-in or trip-out waveform from a disturbance waveform, and if the classification result is the disturbance waveform, the classification result of the disturbance waveform is output. If the classification result is a trip-in or trip-out waveform, a sixth XGBoost binary classifier is used to distinguish a trip-in waveform from a trip-out waveform and output the classification result of the trip-in waveform or the trip-out waveform.

[0014] The application realizes accurate classification of power frequency waveforms of a power transmission line by inputting each calculated waveform feature into a preset cascade classifier model and using multiple XGBoost binary classifiers sequentially cascaded in the model to perform specific binary classification tasks. Specifically, a first XGBoost binary classifier is used to distinguish between a first preset waveform and a second preset waveform. If the result is the second preset waveform, the classification result is directly output. If the result is the first preset waveform, a second XGBoost binary classifier is used to distinguish between the first preset waveform and a non-first preset waveform, and the classification result is output accordingly. For the non-first preset waveform, a third XGBoost binary classifier is used to distinguish between a two-segment waveform and a three-segment waveform. If the result is the two-segment waveform, a fourth XGBoost binary classifier is used to distinguish between an incoming power waveform and a power outage waveform. If the result is the three-segment waveform, a fifth XGBoost binary classifier is used to distinguish between a tripping or string-in waveform and a disturbance waveform. Finally, a sixth XGBoost binary classifier is used to distinguish between a tripping waveform and a string-in waveform. This cascade structure allows each classifier to only handle relatively simple binary classification tasks, reducing the difficulty of classification and improving the accuracy and efficiency of classification. At the same time, the output result of the previous classifier is directly used to trigger the start of the next classifier, forming an efficient classification process that reduces unnecessary calculations and further improves the operating efficiency of the system.

[0015] Further, the cascade classifier model is obtained by training multiple initial XGBoost binary classifiers based on historical power frequency waveform data labeled with preset waveform types and combining the waveform features. Obtain historical power frequency waveform data, classify and label the data according to preset waveform types, calculate the preset waveform features corresponding to the labeled data, and obtain a sample data set. Divide the sample data set into a training set and a test set, and configure training parameters for multiple initial XGBoost binary classifiers. Train each first initial XGBoost binary classifier based on the training set to enable each classifier to learn binary classification logic and obtain each second initial XGBoost binary classifier. Verify the classification performance of each second initial XGBoost binary classifier based on the test set and select classifiers that meet preset standards to obtain each third initial XGBoost binary classifier. Cascade each third initial XGBoost binary classifier according to a preset classification hierarchy to associate the output of a previous classifier with the start condition of a next classifier, and obtain a cascade classifier model.

[0016] The application can derive significant beneficial effects from the model training logic level through the above cascade classifier model construction process. First, the historical power frequency waveform data is obtained and labeled according to the preset waveform type, and the corresponding preset features are calculated to form a sample data set. This process can convert the typical rules of historical waveforms (such as the three-section feature of the tripping waveform and the amplitude mutation feature of the incoming power waveform) into structured data corresponding to the "feature-type" pair, avoiding learning bias caused by unannotated data or ambiguous features during model training, and providing a high-quality data foundation for subsequent classifier accurate learning of binary classification logic. Second, the sample data set is divided into a training set and a test set, which can ensure that the model learns general rules during the training phase and verifies the generalization ability during the verification phase, avoiding the overfitting problem that may occur when training a single data set. The training parameters (such as the number of base learners and the learning rate) of the initial XGBoost binary classifier are configured specifically to balance the model complexity and learning accuracy, so that each classifier can fully learn the feature mapping relationship of a specific binary classification task (such as normal / second preset waveform classification and two-section / three-section waveform classification), and will not be caused by improper parameters. The training efficiency is low or the classification bias; third, after obtaining the second initial XGBoost binary classifier through training, the third initial XGBoost binary classifier that meets the preset standard is selected based on the test set, which can eliminate models with insufficient classification performance and ensure that each classifier participating in the cascade has reliable classification ability, thereby guaranteeing the overall performance of the cascade model from the component level; finally, the selected classifiers are cascaded in order according to the preset classification level relationship, so that the output of the previous level classifier is associated with the starting condition of the next level classifier (such as the second preset waveform classification result directly outputting and not triggering the subsequent classification), which can reduce the task complexity of a single classifier through the "layer-by-layer subdivision" logic (compared with a single multi-classification model, it is easier to master the waveform difference), and can reduce redundant calculations through "invalid path truncation". At the same time, the XGBoost architecture does not need to rely on a deep learning framework, which can reduce the resource consumption of training and running. In summary, the model construction process can make the final cascade classifier model have the advantages of high classification accuracy, strong generalization ability and low resource consumption through the progressive design of "high-quality data preparation-scientific data set division-accurate parameter configuration-performance screening-logical cascade", which can provide reliable model support for accurate classification of power frequency waveforms of transmission lines.

[0017] Further, the preset waveform type includes a first preset waveform, a second preset waveform, an incoming power waveform, a power failure waveform, a tripping waveform, a string-in waveform, and a disturbance waveform, specifically: The power frequency waveform of the line normal operation and regular periodic change is determined as the first preset waveform; The power frequency waveform deviating from the preset normal amplitude range or without periodic characteristics is determined as the second preset waveform; determine the two-section power frequency waveform with the amplitude abruptly changing from the preset low level to the preset normal operation level as the incoming power waveform; determine the two-section power frequency waveform with the amplitude abruptly changing from the preset normal operation level to the preset low level as the power-off waveform; determine the three-section power frequency waveform including the preset normal operation section, the amplitude surge section and the amplitude decay to the preset low level section as the trip waveform; determine the three-section power frequency waveform including the preset low amplitude section, the preset high amplitude section and the recovery to the initial preset amplitude section as the inrush waveform; determine the three-section power frequency waveform including the preset normal operation section, the amplitude decreasing to the preset low level section within the preset time and the recovery to the preset normal operation section as the disturbance waveform.

[0018] The present application provides a clear basis for the classification of power frequency waveforms by defining preset waveform types, including first preset waveforms, second preset waveforms, incoming power waveforms, power-off waveforms, trip waveforms, inrush waveforms and disturbance waveforms. Specifically, the first preset waveforms correspond to power frequency waveforms that are normally operated and regularly periodically changed, and the second preset waveforms correspond to power frequency waveforms that deviate from the preset normal amplitude range or have no periodic characteristics. The incoming power waveform and the power-off waveform correspond to two-section power frequency waveforms with the amplitude abruptly changing from a low level to a normal operation level and from a normal operation level to a low level, respectively. The trip waveform, the inrush waveform and the disturbance waveform correspond to three-section power frequency waveforms containing specific section characteristics, respectively. This clear definition of waveform types enables the classifier to accurately classify based on the characteristics of the waveforms, thereby improving the accuracy and reliability of the classification. In this way, the present application can effectively distinguish different types of power frequency waveforms and provide strong support for fault location of transmission lines.

[0019] In a second aspect, the present application provides a power frequency waveform classification device. The power frequency waveform classification device comprises: an acquisition module configured to acquire preprocessed power frequency waveform data of a transmission line to be classified; a calculation module configured to calculate various waveform features of the power frequency waveform data of the transmission line based on the power frequency waveform data of the transmission line; a classification module configured to input the various waveform features into a preset cascaded classifier model, so that the cascaded classifier model performs a binary classification task according to a plurality of XGBoost binary classifiers internally connected in sequence, and the classification result of a previous XGBoost binary classifier is used to trigger a subsequent XGBoost binary classifier to start or directly output a power frequency waveform classification result; The cascade classifier model is obtained by training a plurality of initial XGBoost binary classifiers based on historical power frequency waveform data labeled with preset waveform types and combining the waveform features, and is sequentially cascaded.

[0020] The power frequency waveform classification device provided by the present application realizes efficient classification of power frequency waveform data of a power transmission line through modular design. Specifically, the acquisition module is responsible for collecting preprocessed power frequency waveform data, ensuring the quality and consistency of the input data; the calculation module calculates a plurality of waveform features based on these data, providing key information for subsequent classification. The classification module inputs these features into a preset cascade classifier model, which is composed of a plurality of XGBoost binary classifiers cascaded in sequence, and each classifier is responsible for performing a specific binary classification task. This cascade structure not only reduces the classification difficulty, but also improves the accuracy and efficiency of classification. The output result of the previous classifier is directly used to trigger the start of the next classifier, forming an efficient classification process. In addition, the cascade classifier model is trained based on historical power frequency waveform data labeled with preset waveform types, and these preset waveform types cover various typical waveforms in the operation of the power transmission line, so that the model can accurately identify and classify different types of power frequency waveforms. Therefore, the power frequency waveform classification device of the present application not only improves the accuracy and efficiency of classification, but also enhances the reliability and practicality of the system, providing strong technical support for fault location of the power transmission line.

[0021] In a third aspect, the present application provides a computer readable storage medium comprising a stored computer program, wherein the computer program, when executed, controls the device in which the computer readable storage medium is located to perform the power frequency waveform classification method as described. The beneficial effects thereof are the same as those of the power frequency waveform classification method provided in the first aspect of the present application.

[0022] In a fourth aspect, the present application provides a terminal device comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements any of the power frequency waveform classification methods described in the first aspect when executing the computer program. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 : an embodiment flowchart of the power frequency waveform classification method provided by the present application; Figure 2 : an embodiment structure diagram of the cascade of six XGBoost classifiers provided by the present application; Figure 3An embodiment structure schematic diagram of a power frequency waveform classification device provided for the application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the application will be apparently and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0025] Embodiment one Please refer to Figure 1 In order to solve the problem that the prior art cannot accurately and efficiently classify power frequency for fault positioning of a power transmission line, an embodiment of the application provides a power frequency waveform classification method, which comprises steps S01-S03.

[0026] S01: obtaining preprocessed power frequency waveform data of a power transmission line to be classified.

[0027] As a preferred embodiment of the present embodiment, the obtaining of the preprocessed power frequency waveform data of the power transmission line to be classified is specifically: First, the distributed fault positioning device deployed along the power transmission line is used to collect original power frequency waveform data to be classified. The distributed fault positioning device usually carries a Rogowski coil or other current sensing element, which can capture the power frequency current signal in the operation process of the line in real time, forming continuous original power frequency waveform data. These original data directly reflect the current change state of the line at present and are the basic data source for subsequent classification.

[0028] Then, the collected original power frequency waveform data is subjected to filtering processing. Since the original data is easily affected by environmental electromagnetic interference, sensor self-noise and other factors during collection, there may be interference information such as waveform burr and amplitude fluctuation. Therefore, a median filtering algorithm is used to process the original data: by setting a fixed size sliding window, the original power frequency waveform data is traversed, the median of all data points in each window is calculated, and the data point at the center position of the window is replaced with the median, so as to eliminate the influence of random noise on the waveform characteristics, obtain the filtered waveform data, and ensure the accuracy of subsequent feature calculation.

[0029] Subsequently, the filtered waveform data is standardized to unify the data format and scale. Due to the differences in sampling rate and sampling duration of different distributed fault locating devices, the number of data points of the filtered waveform data is inconsistent (for example, the sampling rate of some data is 2 kHz and the sampling duration is 1 s, corresponding to 2000 data points; the sampling rate of some data is 4 kHz and the sampling duration is 1 s, corresponding to 4000 data points). The subsequent waveform feature calculation and classification model training all need to be based on waveform data with a unified data amount, so the number of data points needs to be adjusted according to the preset data standard: if the data amount of the filtered waveform data is less than the preset data amount, the data points are supplemented by using a cubic spline interpolation algorithm, the domain of the filtered waveform is divided into multiple continuous subintervals, a cubic polynomial that meets the conditions of passing the end data points and the continuity of the first and second derivatives at the junction of adjacent subintervals is constructed in each subinterval, the amplitude of the interpolation point is calculated based on the polynomial, and the number of data points is supplemented to the preset data amount; if the data amount of the filtered waveform data is greater than the preset data amount, the data is first processed by using a low-pass filter with anti-aliasing function to filter out high-frequency components higher than the target Nyquist frequency (i.e. half of the target sampling rate), to avoid the frequency aliasing phenomenon in the subsequent downsampling process, and then the filtered data points are extracted at a preset interval by using a resampling algorithm to reduce the number of data points to the preset data amount.

[0030] More specifically, the cubic spline interpolation in the embodiment is as follows: for a power frequency waveform with less than 4000 data points, a cubic spline interpolation algorithm is used for interpolation processing. The algorithm can splice a series of simple cubic polynomial functions into a complete smooth curve, and ensure that the splicing nodes are also smooth. Given a power frequency waveform with n (n<4000) data points, the domain is first divided into n-1 continuous subintervals. On each subinterval, a cubic polynomial of the form Si(x) = ai+bi(x-xi)+ci(x-xi)2+di(x-xi)3 is constructed, where ai, bi, ci, di are undetermined coefficients. To ensure that the overall curve after splicing has continuous and smooth characteristics, the algorithm strictly requires that all segment functions satisfy the following constraints: each segment function must pass through the known data points at the interval endpoints, which ensures the accuracy of the interpolation; at all internal nodes, the two adjacent segment functions must have equal first and second derivatives, which ensures that the curve is smooth at the nodes without sharp corners and the curvature smoothly transitions. To close the equation set and obtain a unique solution, natural boundary conditions are usually imposed on the first and last endpoints, i.e., the second derivative is set to zero. By solving the large sparse linear equation set thus formed, the coefficients of all polynomials can be determined. Finally, using the solved cubic polynomials for each interval, the function value at any target interpolation point can be calculated, thereby accurately interpolating the original waveform to 4000 data points. The generated smooth curve can effectively overcome the harshness that may be caused by linear interpolation, and better reproduce the true form of the power frequency waveform.

[0031] The resampling with anti-aliasing filtering in the embodiment is as follows: for a power frequency waveform with more than 4000 data points, a resampling algorithm with anti-aliasing filtering is used for processing. The algorithm first designs a low-pass filter with a cutoff frequency of the target Nyquist frequency (i.e., half of the target sampling rate), and applies this anti-aliasing filter to the original waveform to pre-attenuate or eliminate all frequency components above the target Nyquist frequency. After completing this filtering operation to ensure that the signal spectrum is limited within a safe range, the algorithm performs proportional downsampling to extract data points from the filtered waveform at intervals, and finally reduces the number of waveform data points to 4000. This process effectively compresses the data amount while maximizing the retention of effective frequency domain information of the power frequency waveform at the target resolution, ensuring the authenticity and accuracy of the downsampled waveform.

[0032] After the above filtering and standardization processing, the preprocessed power frequency waveform data of the transmission line to be classified can be obtained. The preprocessed data not only eliminates noise interference, but also unifies the data format, can provide stable and consistent input for subsequent waveform feature calculation, and ensure the accuracy and reliability of the subsequent power frequency waveform classification process.

[0033] S02: based on the power frequency waveform data of the power transmission line, calculating each waveform feature of the power frequency waveform data of the power transmission line.

[0034] As a preferred embodiment of the present embodiment, the calculation of each waveform feature of the power frequency waveform data of the power transmission line based on the power frequency waveform data of the power transmission line is specifically: After obtaining the pre-processed power frequency waveform data of the power transmission line, various pre-set waveform features need to be calculated based on the data. These features can accurately capture the core differences of different types of power frequency waveforms, and provide key basis for subsequent classification. The specific calculation process is as follows: First, calculate the statistical features of the waveform. This kind of feature mainly reflects the overall distribution law of the waveform amplitude. For the pre-processed power frequency waveform data, the average value is obtained by performing arithmetic operation on the amplitude of all data points, that is, the sum of the amplitudes of all data points divided by the total number of data points, which is used to represent the overall level of the waveform amplitude; the standard deviation is calculated, the deviation of each data point amplitude from the average value is first calculated, the sum of the squares of all deviations is then calculated, and the sum is divided by the total number of data points and square root, which is used to reflect the dispersion degree of the waveform amplitude; the amplitude range is calculated, the maximum value and the minimum value of the waveform data are selected, and the maximum value is subtracted from the minimum value, which is used to reflect the fluctuation span of the waveform amplitude.

[0035] Secondly, the frequency energy feature of the waveform is calculated. This kind of feature is used to distinguish the energy proportion of different frequency components in the waveform. First, the frequency domain conversion is performed on the pre-processed power frequency waveform data, and the amplitude signal in time domain is converted into frequency-energy distribution in frequency domain; then, three frequency intervals of high frequency, medium frequency and low frequency are divided, and the absolute values of the amplitudes of all frequency components in each interval are accumulated to obtain high frequency energy, medium frequency energy and low frequency energy, respectively. The high frequency energy corresponds to the energy sum of the high frequency interval, the medium frequency energy corresponds to the energy sum of the medium frequency interval, and the low frequency energy corresponds to the energy sum of the low frequency interval. The three types of energy features can effectively distinguish the second preset waveform containing high frequency interference and the first preset waveform mainly containing low frequency.

[0036] Finally, the shape features of the waveform are calculated, which focus on the amplitude variation law and shape difference of the waveform. When counting the number of wave peaks, set the amplitude threshold value that meets the characteristics of the power frequency waveform, traverse all data points, identify the wave peak whose amplitude is greater than the threshold value and the amplitudes of the adjacent data points on both sides are less than that of the point, and count the identified wave peaks; when calculating the smoothness feature, the original power frequency waveform data before preprocessing is called, the amplitude difference between the corresponding data points of the original waveform and the median filtered waveform is compared, the sum of the absolute values of all differences is calculated, the larger the value, the more serious the noise interference on the original waveform, which can be used to distinguish the second preset waveform from the first preset waveform; when calculating the stationarity feature, the preprocessed waveform data is divided into multiple windows of fixed size, the mean of the data points in each window is calculated, and the standard deviation of all window means is calculated, the smaller the standard deviation, the more stable the waveform amplitude, which can be used to distinguish the steady waveform (first preset steady waveform) from the non-steady waveform (non-first preset steady waveform).

[0037] The average, standard deviation, amplitude range, high-frequency energy, medium-frequency energy, low-frequency energy, smoothness feature, wave peak number and stationarity feature calculated by the above process together constitute a complete feature set for power frequency waveform classification, which can fully cover the core differences of different types of power frequency waveforms and provide sufficient support for accurate classification of subsequent cascade classifier models.

[0038] S03: inputting the various waveform features into a preset cascade classifier model, so that the cascade classifier model performs a binary classification task according to multiple XGBoost binary classifiers connected in sequence inside, and the classification result of a previous XGBoost binary classifier is used to trigger the start of a subsequent XGBoost binary classifier or directly output the power frequency waveform classification result; The cascade classifier model is obtained by sequentially connecting multiple initial XGBoost binary classifiers trained based on historical power frequency waveform data labeled with preset waveform types and the various waveform features; the preset waveform types include a first preset waveform, a second preset waveform, an incoming waveform, a power-off waveform, a tripping waveform, a string-in waveform and a disturbance waveform.

[0039] As a preferred embodiment of the present embodiment, the cascade classifier model is obtained by sequentially connecting multiple XGBoost binary classifiers connected in sequence inside, and the classification result of a previous XGBoost binary classifier is used to trigger the start of a subsequent XGBoost binary classifier or directly output the power frequency waveform classification result, specifically: After the waveform feature calculation of the power frequency waveform data of the power transmission line is completed, the obtained various waveform features are input into a preset cascade classifier model, and a plurality of XGBoost binary classifiers inside the model are sequentially cascaded to cooperatively perform a classification task, and finally the classification result of the power frequency waveform is output. The specific execution process is as follows: First, start the cascade classifier model and load the plurality of XGBoost binary classifiers pre-configured inside. The classifiers are sequentially associated in a preset order to form a "layer-by-layer subdivision" classification logic chain. The complete waveform feature set calculated is taken as the initial input and is transmitted to the first XGBoost binary classifier of the cascade classifier model. The specific task of the classifier is to distinguish between the "first preset waveform" and the "second preset waveform". By analyzing the key indicators such as the amplitude range, high-frequency energy, and smoothness feature in the waveform feature (for example, the second preset waveform often shows that the amplitude deviates from the normal range, the high-frequency energy is too high, or the smoothness feature value is too large), the corresponding classification result is output. If the classification result is "second preset waveform", the result is directly taken as the final power frequency waveform classification result and output without triggering the subsequent classifier; if the classification result is "first preset waveform", a trigger signal is generated to start the second XGBoost binary classifier.

[0040] Then, after receiving the trigger signal, the second XGBoost binary classifier reads the same waveform feature set and performs the task of distinguishing between the "first preset steady-state waveform" and the "non-first preset steady-state waveform". The classifier mainly relies on the smoothness feature, the number of wave peaks, and the frequency energy feature (for example, the first preset steady-state waveform has a small smoothness feature value, a regular number of wave peaks, and a high low-frequency energy proportion, while the non-first preset steady-state waveform often shows poor smoothness, fluctuation in the number of wave peaks, or changes in the proportion of medium and high-frequency energy). The second type of classification result is output. If the result is "first preset steady-state waveform", the result is directly output; if it is "non-first preset steady-state waveform", the third XGBoost binary classifier is triggered to start.

[0041] Subsequently, the third XGBoost binary classifier distinguishes between "two-section waveform" and "three-section waveform". The classification basis mainly includes the change rule of the number of wave peaks, the local window feature, and the amplitude mutation feature at the waveform segmentation position (two-section waveform has one obvious amplitude mutation and corresponds to two groups of regular wave peaks; three-section waveform has two amplitude mutations and corresponds to three groups of regular wave peaks). If the classification result is "two-section waveform", the fourth XGBoost binary classifier is triggered; if it is "three-section waveform", the fifth XGBoost binary classifier is triggered.

[0042] When the fourth XGBoost binary classifier is started, the "incoming waveform" and "power-off waveform" are classified by analyzing the average value difference of the wave crest in different intervals (the incoming waveform shows that the average value of the wave crest in the first half is low, and the average value of the wave crest in the second half is high, and the power-off waveform is the opposite), and the classification result is directly output. The fifth XGBoost binary classifier distinguishes between "trip / insertion waveform" and "disturbance waveform" based on the trend of the average value of the wave crest (the trip / insertion waveform shows a "low-high-low" trend, and the disturbance waveform shows a "high-low-high" trend). If the result is "disturbance waveform", it is directly output; if it is "trip / insertion waveform", the sixth XGBoost binary classifier is triggered.

[0043] Finally, the sixth XGBoost binary classifier distinguishes between "trip waveform" and "insertion waveform" by analyzing the amplitude change characteristics of the waveform at the end (the amplitude of the trip waveform at the end decays rapidly or slowly to a low level, and the amplitude of the insertion waveform at the end recovers to the initial level and maintains a sinusoidal rule), and outputs the final classification result.

[0044] During the entire classification process, the result of the previous XGBoost binary classifier determines whether to trigger the subsequent classifier, and avoids redundant calculation of invalid classification steps. At the same time, each classifier only focuses on a single binary classification task, which can more accurately capture the differences of the corresponding waveform type and ensure the accuracy and reliability of the classification result.

[0045] Further, the setting process of the cascaded classifier model of the present application is as follows: As shown in Figure 2 The classifier 1-classifier 6 in the figure are six XGBoost classifiers, which can divide the power frequency waveform into seven types in total. The explanation of each type of power frequency waveform is as follows: 1) The first preset waveform refers to the power frequency current waveform when the line is not faulty, which is usually a sinusoidal waveform (frequency is 50Hz), and some first preset steady-state waveforms may have small disturbances or burrs.

[0046] 2) The second preset waveform refers to the waveform collected when the sensor is abnormal, which is usually a waveform close to 0 or significantly exceeding the normal power frequency current amplitude.

[0047] 3) The incoming waveform refers to the power frequency current waveform when the power is restored or the load is increased, which is composed of two waveforms with obvious amplitude difference, and each waveform is usually a sinusoidal waveform. The overall amplitude of the first waveform is lower than that of the second waveform.

[0048] 4) The power-off waveform refers to the power frequency current waveform when the line is powered off or the load is reduced, which is similar to the incoming waveform and is also composed of two waveforms. The overall amplitude of the first waveform is higher than that of the second waveform.

[0049] 5) Tripping waveform: refers to the power frequency current waveform of the fault phase when the line trips due to a fault, which consists of three segments. The first segment waveform corresponds to the part before tripping, which is similar to the first preset steady-state waveform; the second segment waveform is the part when the line has a fault but has not yet tripped, at which time the waveform is still close to a sine waveform, but the amplitude increases sharply; the third segment waveform is the part after the line trips, at which time the waveform quickly or slowly decays to a value close to zero.

[0050] 6) Inrush waveform: refers to the power frequency current waveform of the non-fault phase when the line trips due to a fault, or the power frequency waveform caused by a short-term power frequency current rise due to related operations of the line, which consists of three segments. The three segments are close to sine waveforms in shape, and the overall amplitude is "low-high-low".

[0051] 7) Disturbance waveform: refers to the power frequency waveform caused by a short-term current drop due to related operations of the line, which consists of three segments. The three segments are close to sine waveforms in shape, and the overall amplitude is "high-low-high".

[0052] To realize reliable classification of the above power frequency waveforms, each XGBoost classifier is configured as follows: (1) Classifier 1: Classifier 1 is used to determine whether the waveform is the first preset waveform or the second preset waveform (second preset waveform). If it is the first preset waveform, it further enters the subsequent classification process; if it is the second preset waveform, it outputs the second preset waveform classification result.

[0053] For the second preset waveform, the main differences from the first preset waveform are as follows: 1) The waveform amplitude and average value are significantly smaller or significantly larger; 2) The waveform oscillation degree is significantly more intense than the first preset waveform; 3) The waveform has a large high-frequency component.

[0054] In view of the above three aspects, the waveform features used to distinguish the two waveforms are as follows: average value, standard deviation, amplitude range, high-frequency energy (f>150Hz), smoothness feature, first derivative standard deviation and amplitude, second derivative standard deviation and amplitude, zero-crossing point number, peak number, spectral entropy, and stationarity feature.

[0055] Among the above features, high-frequency energy refers to the sum of the absolute values of the frequency component amplitudes greater than 150Hz after fast Fourier transform of the waveform; smoothness feature refers to the sum of the absolute values of the differences between the original waveform and the median filtered waveform at each point; stationarity feature refers to the standard deviation of the mean value of the waveform within each window. For classifier 1, a single window covers 100 waveform sampling points.

[0056] (2) Classifier 2: Classifier 2 is used to determine whether the waveform is the first preset steady-state waveform or a non-first preset steady-state waveform. If it is a non-first preset steady-state waveform, it further enters the subsequent classification process; if it is a first preset steady-state waveform, it outputs the first preset steady-state waveform classification result.

[0057] For the non-first preset steady-state waveform, the main difference from the first preset waveform is reflected in the following two aspects: 1) The waveform is obviously divided into two or more segments, and the overall amplitude of each segment waveform is obviously different; 2) There are relatively obvious high-frequency components at the junction of each segment waveform, and the overall medium and low frequency components of the waveform are less than those of the first preset steady-state waveform.

[0058] In view of the above two aspects, on the basis of classifier 1, two features of medium frequency energy (50Hz<f<150Hz) and low frequency energy (f<50Hz) are added, and the calculation method is similar to that of the aforementioned high frequency energy. The waveform features used to distinguish the two waveforms after adding are as follows: mean, standard deviation, amplitude range, high frequency energy, medium frequency energy, low frequency energy, smoothness feature, first derivative standard deviation and amplitude, second derivative standard deviation and amplitude, zero-crossing point number, wave peak number, spectral entropy, stationarity feature.

[0059] (3) Classifier 3: Classifier 3 is used to determine whether the waveform is a two-segment waveform or a three-segment waveform. If it is a two-segment waveform, it is further judged by classifier 4 whether it is a power-on waveform or a power-off waveform; if it is a three-segment waveform, it is further judged by classifier 5 whether it is a trip / series-in waveform or a disturbance waveform. The main difference between the two-segment waveform and the three-segment waveform is the number of waveform segments with significantly different overall amplitudes. The features used to distinguish the two waveforms are as follows: wave peak statistical features, waveform 0%-10% and 10%-20% region feature difference, waveform 80%-90% and 90%-100% region feature difference, local window feature, wavelet transform feature.

[0060] Among the above features, the wave peak statistical features mainly include standard deviation, average value, amplitude range, amplitude sum, and slope of positive / negative wave peaks, which are used to preliminarily judge the number of waveform segments by wave peaks. The local window features refer to the average value, standard deviation, and amplitude range of all waveforms in a specific window range, which are used to capture the statistical features of the connection areas of each waveform segment. For the classifier 3, a single window covers 800 sampling points. The wavelet transform features refer to the average value and standard deviation of the detail coefficients and approximation coefficients obtained by discrete wavelet transform, which are used to highlight the mutation of the waveform. For the classifier 3, the wavelet transform uses db4 wavelet, and the decomposition level is 1. The difference between the 0%-10% and 10%-20% region features of the waveform and the difference between the 80%-90% and 90%-100% region features of the waveform mainly include the difference between the average value, standard deviation, amplitude range, and smoothness index, which are mainly used to solve the problem that the model is not sensitive enough to the mutation occurring at the beginning and end of the waveform.

[0061] (4) Classifier 4: The classifier 4 is used to distinguish whether the two-segment waveform is an incoming waveform or a power-off waveform. The difference between the two lies in the average value of the wave peaks of the front and rear two segments. The wave peak of the first segment of the incoming waveform is smaller than that of the second segment, and the power-off waveform is the opposite. The features used to distinguish the two waveforms are as follows: the average value of the wave peaks of the 0%-10% and 10%-20% regions of the waveform, and the average value of the wave peaks of the 80%-90% and 90%-100% regions of the waveform.

[0062] (5) Classifier 5: The classifier 5 is used to judge whether the waveform is a trip / inrush waveform or a disturbance waveform. If it is a trip / inrush waveform, it further enters the subsequent classification process; if it is a disturbance waveform, it outputs the classification result of the disturbance waveform.

[0063] The main difference between the trip / inrush waveform and the disturbance waveform lies in the trend of the average value of the wave peaks of the three segments. The average value of the wave peaks of the trip / inrush waveform is low on both sides and high in the middle; the average value of the wave peaks of the disturbance waveform is high on both sides and low in the middle. Since the positions of the connection areas of the three-segment waveforms are not fixed, it is difficult to accurately classify the two types of waveforms by simply using the average value of the wave peaks as in the classifier 4, so the features used to distinguish the two types of waveforms are as follows: wave peak statistical features, local window features, and wavelet transform features. The number of sampling points covered by the window of the classifier 5 and the wavelet transform parameters are the same as those of the classifier 3.

[0064] (6) Classifier 6: The classifier 6 is used to determine whether the waveform is a trip waveform or an inrush waveform. The average values of the peaks of each segment of the two types of waveforms are low on both sides and high in the middle. The difference between the two is the shape and amplitude of the third segment of the wave. For the trip waveform, the third segment of the waveform should quickly decay to 0 from the specified time or slowly decay to a value close to 0 according to the exponential function. For the inrush waveform, the shape of the third segment of the wave is still close to the sine waveform and will not decay to 0. Considering the above characteristics, the features used to distinguish the two types of waveforms are as follows: peak statistical features, 80%-90% and 90%-100% region feature difference, local window feature, and wavelet transform feature. The window coverage of the sampling points and the wavelet transform parameters of the classifier 6 are the same as those of the classifier 3.

[0065] Parameter setting before training: Set the proportion of training set and test set to 80% and 20% respectively, and confirm some training parameters in the training process, mainly including n_estimators, max_depth, learning_rate, subsample, colsample_bytree, objective and other parameters. The specific parameter description and corresponding values are as follows: (1) n_estimators: the number of base learners (trees), the larger the value, the more complex the model, and too large a value may cause the classifier to overfit. For the power frequency waveform classification task containing 4000 sampling points, the value can be set to 1000.

[0066] (2) max_depth: the maximum depth of a single tree, used to control the complexity of the model, which can be set to 10 to reduce overfitting as much as possible while ensuring classification accuracy.

[0067] (3) learning_rate: learning rate (step size), used to control the contribution weight of each tree to the final result, which can be set to 0.01 to capture the weak feature differences of each power frequency waveform.

[0068] (4) subsample: the proportion of randomly sampled samples during the training of each tree, which can be set to 0.5 to effectively prevent overfitting and enhance data diversity.

[0069] (5) colsample_bytree: the proportion of randomly sampled features for each tree, which can be set to 0.8 to retain most of the key features.

[0070] (6) objective: used to determine whether the model is used for binary classification or multi-classification task, which can be set to 'binary:logistic', i.e. the model is used for binary classification task, and the output probability value.

[0071] In summary, the embodiment can eliminate noise interference and format differences (such as uniform data volume) of the original data by obtaining the preprocessed power frequency waveform data of the power transmission line, provide consistent input for subsequent feature calculation, avoid classification deviation caused by irregular data, and lay the foundation for accurate classification. Secondly, based on data, the various waveform features (such as average value, frequency domain energy, and wave peak number) of the power frequency waveform data of the power transmission line are calculated. These features are designed for the core differences of different waveform types (such as amplitude anomaly of the second preset waveform and three-section amplitude change of the tripping waveform). Without model self-exploration features, the key distinguishing dimensions can be directly focused on, the invalid learning cost is reduced, the feature interpretability is improved, and the classification accuracy is improved. Furthermore, the feature input is input into the model cascaded by multiple XGBoost binary classifiers. Each classifier only performs a specific binary classification task (such as distinguishing normal / second preset waveform and then subdividing steady-state / non-steady-state waveform). Compared with the traditional single multi-classification model, the task complexity of a single classifier is greatly reduced, the model is more likely to learn the subtle differences between various waveforms, and the logic of the previous stage classification result triggers the next stage classification (such as the second preset waveform is directly output without subsequent classification). Redundant calculation is reduced, and classification efficiency is improved. At the same time, the XGBoost model does not need to rely on a deep learning framework and has no special requirements for GPU computing power, which can significantly reduce resource consumption in the training and running process and avoid performance bottlenecks. Finally, the cascaded classifier is trained based on historical data labeled with preset waveform types and preset features, which can fully utilize the typical rules of historical waveforms, so that the model can still be stable when facing complex waveforms such as tripping and stringing, and the generalization ability is improved. In summary, through the technical logic of "standardized data-clear features-staged classification-lightweight architecture", the application can simultaneously improve the accuracy, efficiency and generalization ability of power frequency waveform classification while reducing resource consumption, providing reliable support for automatic diagnosis of power transmission line fault positioning, and further optimizing the efficiency of power grid fault disposal. The application effectively solves the problem that the prior art cannot accurately and efficiently classify the power frequency of the power transmission line fault positioning.

[0072] Embodiment 2 Please refer to Figure 3 The power frequency waveform classification device provided by the embodiment of the application.

[0073] In the embodiment, the power frequency waveform classification device comprises an acquisition module 10, a calculation module 20 and a classification module 30.

[0074] The acquisition module 10 is configured to acquire preprocessed power frequency waveform data of a power transmission line to be classified. The calculation module 20 is configured to calculate various waveform features of the power frequency waveform data of the power transmission line based on the power frequency waveform data of the power transmission line. The classification module 30 is configured to input the waveform features into a preset cascade classifier model, so that the cascade classifier model performs a binary classification task according to a plurality of XGBoost binary classifiers which are sequentially cascaded internally, and a classification result of a previous XGBoost binary classifier is used to trigger a subsequent XGBoost binary classifier to start or directly output a power frequency waveform classification result. The cascade classifier model is obtained by sequentially cascading a plurality of initial XGBoost binary classifiers which are trained based on historical power frequency waveform data labeled with preset waveform types and the waveform features.

[0075] For the convenience and brevity of description, the device embodiments of the present application include all the implementation manners of the power frequency waveform classification method embodiments, which will not be described herein.

[0076] Embodiment three The embodiment of the present application provides a computer readable storage medium, which comprises a stored computer program, wherein the computer program controls a device where the computer readable storage medium is located to execute the power frequency waveform classification method when the computer program is running. The power frequency waveform classification method can be stored in a computer readable storage medium if it is realized in the form of a software function unit and used as an independent product. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can implement the steps of each method embodiment when executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable medium can include any entity or device, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. that can carry the computer program code.

[0077] Embodiment four The embodiment provides a terminal device, which comprises a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, and the processor implements any one of the power frequency waveform classification methods according to the embodiment one when executing the computer program.

[0078] The above-described specific embodiments further illustrate the objects, technical solutions and advantages of the present application. It should be understood that the above-described specific embodiments are merely for the purpose of illustrating the present application and are not intended to limit the scope of protection of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A power frequency waveform classification method, characterized by, The method comprises the following steps: acquiring preprocessed power frequency waveform data of a power transmission line to be classified; calculating various waveform features of the power frequency waveform data of the power transmission line based on the power frequency waveform data of the power transmission line; inputting the various waveform features into a preset cascade classifier model, so that the cascade classifier model performs a binary classification task according to a plurality of XGBoost binary classifiers which are sequentially cascaded inside, and the classification result of a previous XGBoost binary classifier is used to trigger the start of a subsequent XGBoost binary classifier or directly output a power frequency waveform classification result; wherein the cascade classifier model is obtained by sequentially cascading a plurality of initial XGBoost binary classifiers which are trained based on historical power frequency waveform data labeled with preset waveform types and the various waveform features; and the preset waveform types include a first preset waveform, a second preset waveform, an incoming power waveform, a power failure waveform, a tripping waveform, a string-in waveform and a disturbance waveform.

2. The power frequency waveform classification method of claim 1, wherein The acquisition of the preprocessed power frequency waveform data of the power transmission line to be classified is specifically as follows: acquiring original power frequency waveform data to be classified collected by a distributed fault location device of the power transmission line; performing median filtering processing on the original power frequency waveform data to be classified to obtain filtered waveform data; if the data amount of the filtered waveform data is less than a preset data amount, data points are supplemented through interpolation; and if the data amount of the filtered waveform data is greater than the preset data amount, data points are reduced through decimation, so that the filtered waveform data meets a preset data standard, and the preprocessed power frequency waveform data of the power transmission line to be classified is obtained.

3. The power frequency waveform classification method of claim 1, wherein The calculation of the various waveform features of the power frequency waveform data of the power transmission line based on the power frequency waveform data of the power transmission line is specifically as follows: the various waveform features include an average value, a standard deviation, an amplitude range, high-frequency energy, medium-frequency energy, low-frequency energy, a smoothness feature, a number of wave peaks and a stationarity feature; the average value, the standard deviation and the amplitude range are obtained by performing arithmetic operations on the amplitudes of the power frequency waveform data of the power transmission line; the high-frequency energy, the medium-frequency energy and the low-frequency energy are obtained by performing frequency domain conversion on the power frequency waveform data of the power transmission line, dividing different frequency intervals and accumulating energy components of each frequency interval; the number of wave peaks is obtained based on the amplitude variation law of the power frequency waveform data of the power transmission line; the smoothness feature is calculated based on the difference between the original power frequency waveform data to be classified and the power frequency waveform data of the power transmission line, and the stationarity feature is obtained by dividing data windows and analyzing the stability of data in the windows.

4. The power frequency waveform classification method of claim 1, wherein The inputting of the various waveform features into the preset cascade classifier model so that the cascade classifier model performs a binary classification task according to a plurality of XGBoost binary classifiers which are sequentially cascaded inside, and the classification result of a previous XGBoost binary classifier is used to trigger the start of a subsequent XGBoost binary classifier or directly output a power frequency waveform classification result is specifically as follows: the various waveform features are input into the preset cascade classifier model, so that the cascade classifier model first distinguishes the first preset waveform from the second preset waveform through the first XGBoost binary classifier, and outputs the classification result of the second preset waveform if the classification result is the second preset waveform; If the classification result is the first preset waveform, a second XGBoost binary classifier is used to distinguish the first preset steady-state waveform from a non-first preset steady-state waveform, and if the classification result is the first preset steady-state waveform, a first preset steady-state waveform classification result is output; If the classification result is a non-first preset steady-state waveform, a third XGBoost binary classifier is used to distinguish a two-segment waveform from a three-segment waveform; If it is a two-segment waveform, a fourth XGBoost binary classifier is used to distinguish a power-on waveform from a power-off waveform, and a power-on waveform or power-off waveform classification result is output; If it is a three-segment waveform, a fifth XGBoost binary classifier is used to distinguish a trip-in or trip-out waveform from a disturbance waveform, and if the classification result is a disturbance waveform, a disturbance waveform classification result is output; If the classification result is a trip-in or trip-out waveform, a sixth XGBoost binary classifier is used to distinguish a trip-in waveform from a trip-out waveform, and a trip-in waveform or trip-out waveform classification result is output.

5. The power frequency waveform classification method of claim 1, wherein, The cascade classifier model is obtained by sequentially cascading a plurality of initial XGBoost binary classifiers trained based on historical power frequency waveform data labeled with preset waveform types and in combination with the respective waveform features, and specifically comprises the following steps: Obtain historical power frequency waveform data, classify and label the data according to preset waveform types, calculate the preset waveform features corresponding to the labeled data, and obtain a sample data set; Divide the sample data set into a training set and a test set, and configure training parameters for a plurality of initial XGBoost binary classifiers; Train each first initial XGBoost binary classifier based on the training set, so that each classifier learns binary classification logic, and obtain each second initial XGBoost binary classifier; Verify the classification performance of each second initial XGBoost binary classifier based on the test set, and select classifiers meeting preset standards to obtain each third initial XGBoost binary classifier; Cascade each third initial XGBoost binary classifier in a preset classification hierarchy, so that the output of the previous stage classifier is associated with the start condition of the next stage classifier, and obtain a cascade classifier model.

6. The power frequency waveform classification method of claim 1, wherein The preset waveform types include a first preset waveform, a second preset waveform, a power-on waveform, a power-off waveform, a trip-in waveform, a trip-out waveform, and a disturbance waveform, and specifically comprise the following steps: Determine a power frequency waveform with normal line operation and regular periodic variation as the first preset waveform; Determine a power frequency waveform deviating from a preset normal amplitude range or without periodic characteristics as the second preset waveform; Determine a two-segment power frequency waveform with a sudden change in amplitude from a preset low level to a preset normal operation level as the power-on waveform; Determine a two-segment power frequency waveform with a sudden change in amplitude from a preset normal operation level to a preset low level as the power-off waveform; Determine a three-segment power frequency waveform including a preset normal operation segment, an amplitude surge segment, and an amplitude decay to a preset low level segment as the trip-in waveform; Determine a three-segment power frequency waveform including a preset low amplitude segment, a preset high amplitude segment, and a recovery to an initial preset amplitude segment as the trip-out waveform; Determine a three-segment power frequency waveform including a preset normal operation segment, an amplitude decrease to a preset low level segment within a preset time, and a recovery to a preset normal operation segment as the disturbance waveform.

7. A power frequency waveform classification device, characterized by comprising: The method comprises the following steps: An acquisition module is configured to acquire preprocessed power frequency waveform data of a power transmission line to be classified; A calculation module is configured to calculate various waveform features of the power frequency waveform data of the power transmission line based on the power frequency waveform data of the power transmission line; A classification module is configured to input the various waveform features into a preset cascade classifier model, so that the cascade classifier model performs a binary classification task according to a plurality of XGBoost binary classifiers which are sequentially cascaded inside the cascade classifier model, and a classification result of a previous XGBoost binary classifier is used to trigger a subsequent XGBoost binary classifier to start or directly output a power frequency waveform classification result. The cascade classifier model is obtained by sequentially cascading a plurality of initial XGBoost binary classifiers which are trained based on historical power frequency waveform data labeled with preset waveform types and the various waveform features; and the preset waveform types include a first preset waveform, a second preset waveform, an incoming power waveform, a power-off waveform, a tripping waveform, a string-in waveform, and a disturbance waveform.

8. A power frequency waveform classification device according to claim 7, characterised in that, The inputting of the various waveform features into the preset cascade classifier model, so that the cascade classifier model performs the binary classification task according to the plurality of XGBoost binary classifiers which are sequentially cascaded inside the cascade classifier model, and the classification result of the previous XGBoost binary classifier is used to trigger the subsequent XGBoost binary classifier to start or directly output the power frequency waveform classification result, specifically includes: The inputting of the various waveform features into the preset cascade classifier model, so that the cascade classifier model first distinguishes the first preset waveform from the second preset waveform through a first XGBoost binary classifier, and outputs a classification result of the second preset waveform if the classification result is the second preset waveform; If the classification result is the first preset waveform, the cascade classifier model distinguishes a first preset steady-state waveform from a non-first preset steady-state waveform through a second XGBoost binary classifier, and outputs a classification result of the first preset steady-state waveform if the classification result is the first preset steady-state waveform; If the classification result is the non-first preset steady-state waveform, the cascade classifier model distinguishes a two-segment waveform from a three-segment waveform through a third XGBoost binary classifier; If the waveform is the two-segment waveform, the cascade classifier model distinguishes the incoming power waveform from the power-off waveform through a fourth XGBoost binary classifier, and outputs a classification result of the incoming power waveform or the power-off waveform; If the waveform is the three-segment waveform, the cascade classifier model distinguishes a tripping or string-in waveform from a disturbance waveform through a fifth XGBoost binary classifier, and outputs a classification result of the disturbance waveform if the classification result is the disturbance waveform; If the classification result is the tripping or string-in waveform, the cascade classifier model distinguishes the tripping waveform from the string-in waveform through a sixth XGBoost binary classifier, and outputs a classification result of the tripping waveform or the string-in waveform.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium includes a stored computer program, wherein the computer program controls the device where the computer readable storage medium is located to perform the power frequency waveform classification method according to any one of claims 1 to 6 when the computer program is running.

10. A terminal device, comprising: The device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and the processor performs the power frequency waveform classification method according to any one of claims 1 to 6 when the computer program is executed.

Images