Multi-modal power data analysis method, system and equipment based on AI large model

Abstract

The invention discloses a multi-modal power data analysis method, system and device based on an AI large model, and relates to the field of data analysis, and the method comprises the steps: S1, recognizing an abnormal event, and carrying out the first adjustment of a proportion constraint factor; s2, obtaining a fault node, and judging whether a diffusive fault is caused or not; s3, if a diffusive fault is caused, adjusting the monitoring period and the data extraction frequency of each fault node; and S4, if the diffusive fault is not caused, obtaining a second monitoring adjustment parameter of the single fault node and adjusting the monitoring period of the single fault node. Through modal decomposition and proportion constraint factor adjustment, secondary weight adjustment is adopted, overlarge proportion of single modal data is avoided, it is ensured that abnormal data cannot affect overall analysis, fault node positioning is carried out in combination with an impedance method, the judgment accuracy of different types of faults is improved, meanwhile, whether the faults are diffused or not is judged, misjudgment and missed judgment are reduced, and the fault diagnosis accuracy is improved. The period and the data extraction frequency are monitored, and resource allocation is optimized.

Description

Technical Field

[0001] The present invention relates to the field of data analysis, and specifically to a multimodal power data analysis method, system, and device based on an AI large model. Background Art

[0002] Modern power grids contain massive amounts of heterogeneous data from devices such as PMUs (synchronized phasor measurement units), SCADA systems (supervisory control and data acquisition), and smart meters. This data covers multiple modalities, including voltage, current, frequency, power, and harmonics. Traditional methods struggle to effectively integrate and analyze this data at varying frequencies, temporal, and spatial scales. However, large AI models can achieve unified modeling through deep learning methods, enabling more comprehensive analysis.

[0003] Existing technology, such as the invention patent with announcement number CN112256782B, is a Hadoop-based power big data processing system, which relates to the field of big data processing technology. This system uses a data integration subsystem to collect multi-source heterogeneous power data from different data sources, and uses ETL tools to extract, clean, transform, and load data, achieving centralized data collection from different data sources. The data is then output to a data warehouse subsystem in a fixed format. The data warehouse subsystem stores and consolidates the power data in a file layer through the data warehouse, using the file read / write engine HDFS for data reading and writing support, supporting parallel, multi-layer data processing during big data processing. The data quality management subsystem monitors and manages the power data stored in the data warehouse subsystem after being processed by the ETL tool using configured rules, and submits data quality reports.

[0004] Existing technologies, such as the invention patent with announcement number CN116166857B, are a power data processing system based on big data, including a data acquisition module, which is connected to the Internet to obtain the daily power consumption of each group within the rated time of the corresponding area; a data transmission module, which receives the daily power consumption of each group and transmits the corresponding data to the data processing module; a data processing module, which receives the daily power consumption data of each group and processes the data to obtain the daily standard power consumption; a power allocation module, which obtains the daily standard power consumption and obtains the total power distribution ratio of the current corresponding area based on the daily standard power consumption. The power grid dispatching room distributes power to each area according to the current distribution ratio, and corrects the data of each group of daily power consumption through one round of screening preprocessing and two rounds of feature extraction processing.

[0005] Based on the above scheme, it can be seen that in the field of power data analysis, existing technologies often rely on preset rules for data quality management and analysis. However, the anomalies of power data are complex and dynamically changing, and the nodes of the power network affect each other. It is difficult to adapt to all scenarios only through static rules. In addition, there are many types of power failures. Relying solely on traditional data correction methods is inefficient, and power system failures are generally sudden failures. Therefore, real-time monitoring of the power system is required to prevent emergencies. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the present invention provides a multimodal power data analysis method, system and device based on an AI large model. To achieve the above objectives, the present invention is implemented through the following technical solutions: A multimodal power data analysis method based on an AI large model, comprising:

[0007] S1. Input the real-time data stream in the power grid into the AI big model, monitor the high-frequency modes in the power grid in real time and identify abnormal events, and make the first adjustment to the proportional constraint factor of each modal power data based on the data quality of each modal power data in the abnormal data stream.

[0008] S2. Send the abnormal data stream to the data processing center to obtain the fault node, extract the modal power data of each adjacent node of the fault node for analysis, and determine whether it causes a diffuse fault.

[0009] S3. If a diffusion fault occurs, the hardware performance of the built-in data processing module of each fault node is tested based on the data volume of each modal power data of each fault node. If the hardware performance test result is qualified, the modal power data after adjusting the proportional constraint factor is input into the local processor under the built-in data processing module of each fault node for analysis to obtain the first monitoring adjustment parameter of each fault node for adjustment.

[0010] S4. If no diffusion fault is caused, the faulty device is identified, the data mode to which the faulty device belongs is obtained, the number of faulty devices in each mode of power data is counted, the proportional constraint factor of each mode of power data is adjusted for a second time, and the data is imported into the data processing center for analysis and processing to obtain the second monitoring adjustment parameter of the single fault node and adjust the monitoring period of the single fault node.

[0011] As a preferred technical solution, the proportional constraint factor of each modal power data is adjusted for the first time based on the data quality of each modal power data in the abnormal data stream. The specific process is:

[0012] The real-time data stream of the power grid is obtained through the synchronized phasor measurement unit of the PMU, including the real-time voltage, real-time current, real-time phase and real-time frequency of the power grid.

[0013] The real-time data stream of the power grid is input into the AI large model for time-frequency analysis to obtain the spectrum of the real-time data stream. The spectrum is modally decomposed to separate the high-frequency components, and features are extracted from the high-frequency components. Abnormal events are identified based on the extracted features.

[0014] The start and end time nodes of the abnormal event are extracted and recorded as the abnormal period. The real-time data stream within the abnormal period is intercepted and recorded as the abnormal data stream.

[0015] Modal decomposition of the abnormal data stream is performed to obtain several intrinsic mode functions, which are clustered to obtain high-frequency power data, medium-frequency power data and low-frequency power data, which are recorded as various modal power data.

[0016] The quality detection of each modal power data is performed to obtain the quality detection data of each modal power data, including the noise power spectrum density, the amount of missing data and the amount of outlier data of each modal power data. The quality detection data of each modal power data are respectively averaged to obtain the mean of the quality detection data, including the mean of the noise power spectrum density, the mean of the amount of missing data and the mean of the amount of outlier data. Based on the deviation value of the quality detection data of each modal power data and the mean of the quality detection data, the quality detection value of each modal power data is obtained after correction coupling.

[0017] Based on the quality detection value of each modal power data, a mapping match is performed with the proportional constraint factor corresponding to the interval of each quality detection value pre-stored in the data processing center to obtain the proportional constraint factor of each modal power data and perform the first adjustment.

[0018] As a preferred technical solution, the abnormal data stream is sent to the data processing center to obtain the fault node. The specific process is as follows:

[0019] The abnormal data stream is sent to the data processing center, which extracts abnormal current data and abnormal voltage data from it and determines the abnormal type.

[0020] Based on the abnormal current data and abnormal voltage data, combined with the abnormal type, the fault node is obtained by impedance analysis.

[0021] Obtaining the faulty node specifically includes:

[0022] Based on the real-time voltage and real-time current of each power node, the impedance of each power node is processed and compared with the rated impedance of each power node pre-stored in the built-in database of the data processing center to obtain the impedance deviation value. If the impedance deviation value of a power node is greater than or equal to the impedance deviation threshold, the node is determined to be a faulty node.

[0023] As a preferred technical solution, the extraction of each modal power data of each adjacent node of the faulty node for analysis to determine whether a diffuse fault has occurred specifically includes:

[0024] The real-time data streams of the nodes adjacent to the faulty node are extracted, and the normal data streams of the nodes adjacent to the faulty node are retrieved from the built-in database of the data processing center. After comparison and coupling, the diagnostic deviation value is obtained. Based on the analysis and processing of the diagnostic deviation value, the fault diagnosis results of the nodes adjacent to the faulty node are obtained:

[0025] The fault diagnosis results of the adjacent nodes of the faulty node include node fault and node normal.

[0026] When the diagnosis deviation value of a certain adjacent node is greater than or equal to the diagnosis deviation threshold, the fault diagnosis result of the adjacent node is determined to be a node fault.

[0027] When the diagnosis deviation value of a certain adjacent node is less than the diagnosis deviation threshold, the fault diagnosis result of the adjacent node is determined to be normal.

[0028] If the fault diagnosis result of an adjacent node of the faulty node is a node fault, it is determined that the faulty node causes a diffuse fault.

[0029] The faulty node and the adjacent nodes diagnosed as node faults are collectively referred to as faulty nodes.

[0030] If the fault diagnosis result of an adjacent node of the faulty node is that the node is normal, it is determined that the faulty node has not caused a diffuse fault and is recorded as a single faulty node.

[0031] As a preferred technical solution, if a diffuse fault occurs, hardware performance testing of the built-in data processing module of each fault node is performed based on the data volume of each modal power data of each fault node, specifically including:

[0032] The power data of each mode of each fault node during the abnormal period is retrieved, and after local pre-processing, statistics are performed to obtain the data volume of the power data of each mode of each fault node.

[0033] Obtain hardware performance test data of the built-in data processing module of each faulty node, including the number of CPU threads, CPU instruction execution rate, memory and HDD read and write rates of the built-in data processing module of each faulty node.

[0034] Based on the data volume of each modal power data of each fault node, a mapping and matching is performed with the hardware performance detection verification set corresponding to the data volume interval of each modal power data pre-stored in the built-in database of the data processing center to obtain the hardware performance detection verification set of each fault node, including the number of CPU thread verification, CPU instruction execution verification rate, verification memory and HDD read and write verification rate.

[0035] The hardware performance test data of the built-in data processing module of each fault node is compared and analyzed with the hardware performance test verification set of each fault node to obtain the hardware performance test result of the built-in data processing module of each fault node.

[0036] If the hardware performance test result of the built-in data processing module of a faulty node is qualified, the power data of each mode of the faulty node will be analyzed and processed.

[0037] If the hardware performance test result of the built-in data processing module of a faulty node is unqualified, the demand processing data of the faulty node is imported into the data processing center for analysis and processing.

[0038] As a preferred technical solution, the first monitoring adjustment parameter of each fault node obtained by analysis is adjusted, specifically including:

[0039] The power data of each mode of each fault node is imported into the local processor for processing and analysis. The specific processing process includes:

[0040] Based on the proportional constraint factor after the first adjustment, the local processor adjusts the adoption ratio of each modal power data, and extracts abnormal current data and abnormal voltage data therefrom, determines the abnormal type of each fault node, and based on the fault type of each fault node, matches it with the fault factors corresponding to each fault type pre-stored in the local processor to obtain the fault factor of each fault node, which is used to characterize the severity of the fault type.

[0041] Based on the fault factor of each fault node, a first monitoring adjustment parameter corresponding to the fault factor is extracted, and the first monitoring adjustment parameter of each fault node is obtained to adjust the monitoring period and data extraction frequency of each fault node.

[0042] The first monitoring adjustment parameter of each fault node includes a monitoring period length adjustment parameter and a data extraction frequency adjustment parameter.

[0043] As a preferred technical solution, if no diffuse fault occurs, the faulty device is identified, the data mode to which the faulty device belongs is obtained, the number of faulty devices in each mode of power data is counted, and the proportional constraint factor of each mode of power data is adjusted a second time, specifically including:

[0044] Collect statistics on the real-time performance status parameters and historical usage records of each device at a single fault node.

[0045] The real-time performance status parameters include real-time load current, real-time vibration signal and harmonic content.

[0046] The historical usage records include the cumulative length of time in use, the number of faults and the number of fault types. Real-time performance status verification parameters and historical usage record verification parameters are extracted from the built-in database of the data processing center.

[0047] The real-time performance status verification parameters include the transformer real-time load verification current, the real-time verification vibration signal and the verification harmonic content.

[0048] The historical usage record verification parameters include a verification factor for the cumulative length of time in use, a number of fault verifications, and a number of fault type verifications.

[0049] Based on the real-time performance status parameters and historical usage records of each device, the load index of each device is obtained after comparison and correction coupling with the real-time performance status verification parameters and historical usage record verification parameters. The load index of each device is mapped and matched with the sinusoidal excitation current size corresponding to the load index in the built-in database of the data processing center to obtain the sinusoidal excitation current size of each device.

[0050] A sinusoidal excitation current is generated for each device, and the faulty device is identified based on the response data from each device.

[0051] Perform modal separation and feature extraction on the output data of each faulty device to obtain the main output mode of each faulty device, and count the number of faulty devices in each modal power data.

[0052] Based on the ratio of the number of faulty devices in each modal power data to the total number of faulty devices, a second adjustment is made to the proportional constraint factor of each modal power data in the abnormal period.

[0053] As a preferred technical solution, the data is imported into the data processing center for analysis and processing to obtain the second monitoring adjustment parameter of the single fault node and adjust the monitoring period of the single fault node. The specific processing conditions are:

[0054] Each modal power data is imported into the data processing center. Based on the secondary adjustment proportion constraint factor, the adoption ratio of each modal power data is adjusted. After feature extraction of each modal power data, the fault severity assessment value of the single fault node is comprehensively evaluated. Based on the fault severity assessment value of the single fault node, the fault severity assessment value is mapped and matched with the monitoring period adjustment parameters corresponding to the interval of each fault severity assessment value pre-stored in the built-in database of the data processing center to obtain the second monitoring period adjustment parameter of the single fault node and monitor the single fault node.

[0055] The second monitoring period adjustment parameter of the single fault node includes a monitoring period length adjustment parameter and a data extraction frequency adjustment parameter.

[0056] In addition, a multimodal power data analysis system based on AI large model specifically includes:

[0057] The high-frequency modal monitoring module is used to input the real-time data stream in the power grid into the AI large model, monitor the high-frequency modes in the power grid in real time and identify abnormal events, and make the first adjustment to the proportional constraint factor of each modal power data based on the data quality of each modal power data in the abnormal data stream.

[0058] The diffusion fault judgment module is used to send the abnormal data stream to the data processing center to obtain the fault node, extract the modal power data of each adjacent node of the fault node for analysis, obtain the fault diagnosis results of each adjacent node and determine whether a diffusion fault is caused.

[0059] The multi-node fault processing module is used to perform hardware performance testing on the built-in data processing module of each fault node based on the data volume of each modal power data of each fault node if a diffusion fault occurs. If the hardware performance test result is qualified, the modal power data after adjusting the proportional constraint factor is input into the local processor under the built-in data processing module of each fault node for analysis to obtain the first monitoring adjustment parameter of each fault node and adjust the monitoring period and data extraction frequency of each fault node.

[0060] The single-node fault processing module is used to identify the faulty device if no diffusion fault occurs, obtain the data mode to which the faulty device belongs, count the number of faulty devices in each mode of power data, make a second adjustment to the proportional constraint factor of each mode of power data, and import it into the data processing center for analysis and processing to obtain the second monitoring adjustment parameter of the single fault node and adjust the monitoring period of the single fault node.

[0061] In addition, a multimodal power data analysis device based on an AI large model is provided, wherein the device has one or more programs, and the one or more programs are executed by one or more processors to implement the above method.

[0062] Compared with the prior art, the embodiments of the present invention have at least the following beneficial effects:

[0063] (1) The present invention provides a multimodal power data analysis method based on an AI large model. By performing modal decomposition on the power data and adjusting the proportional constraint factors, the method fully utilizes the data of different frequency ranges of the power grid, making the data analysis more comprehensive and accurate. A secondary weight adjustment is performed on each modal data to avoid excessive weighting of a single modal. When an abnormality occurs in the output device of a certain modal data, the weight of the modal data in the model is reduced, thus avoiding the occurrence of a large amount of erroneous data during the data processing process and improving the accuracy and efficiency of data processing.

[0064] (2) The present invention improves the accuracy of judging different types of faults by locating the fault node in combination with the impedance method. At the same time, it detects diffuse faults, reduces misjudgments and missed judgments, improves diagnostic efficiency, and enhances the safety and reliability of the system. When a diffuse fault occurs, the processing capability of the faulty node is evaluated. For faulty nodes with strong computing power, they are directly processed locally to reduce the computing pressure in the cloud and improve the response speed. For nodes with insufficient computing power, a data processing center is used to ensure the consistency and stability of the analysis. Avoid relying entirely on a general large model, thereby improving computing efficiency and reducing the delay caused by data transmission.

[0065] (3) The present invention analyzes the severity of fault nodes and adjusts the monitoring cycle and data extraction frequency to achieve focused monitoring and improve operation and maintenance efficiency. For nodes with high severity, the monitoring cycle is shortened and the data extraction frequency is increased to avoid unnecessary data processing burdens. More frequent data collection is implemented for seriously faulty nodes, enabling the model to capture more detailed data change trends and identify potential risks in advance, thereby improving the accuracy of early warnings.

[0066] Of course, any product implementing the present invention does not necessarily need to achieve all of the above advantages at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 Schematic diagram of the method of the present invention.

[0068] Figure 2 Schematic diagram of the system module of the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0070] In the description of the present invention, it should be understood that the terms "opening", "upper", "lower", "thickness", "top", "middle", "length", "inside", "around" and the like indicating orientation or positional relationship are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the components or elements referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0071] See also Figure 1 As shown, an embodiment of the present invention provides a multimodal power data analysis method based on an AI large model, specifically comprising:

[0072] S1. Input the real-time data stream in the power grid into the AI big model, monitor the high-frequency modes in the power grid in real time and identify abnormal events, and make the first adjustment to the proportional constraint factor of each modal power data based on the data quality of each modal power data in the abnormal data stream.

[0073] The specific process of S1 is as follows:

[0074] The real-time data stream of the power grid is obtained through the synchronized phasor measurement unit of the PMU, including the real-time voltage, real-time current, real-time phase and real-time frequency of the power grid.

[0075] The real-time data stream of the power grid is input into the AI large model for time-frequency analysis to obtain the spectrum of the real-time data stream. The spectrum is modally decomposed using wavelet transform to separate the high-frequency components, and features are extracted from the high-frequency components. Abnormal events are identified based on the extracted features.

[0076] It should be noted that the wavelet transform is a time-frequency analysis method that can analyze signals at different time scales and is suitable for non-stationary signals, transient signals, harmonic analysis, and mutation signal detection. Wavelet functions are used to perform multi-scale decomposition on real-time data streams to extract characteristic information of different frequency components. Its expression is: Wherein, W(a,b) is the coefficient of wavelet transform, which represents the signal characteristic intensity at scale a and time position b; x(t) is the original signal, i.e., the time series data to be subjected to wavelet transform, which is a real-time data stream in the embodiment of the present invention. is the scaled and translated complex conjugate of the mother wavelet, is the complex conjugate of the wavelet basis function, a is the scale factor (Scale), which is used to control the stretching or compression of the wavelet. a>1 means that the wavelet is stretched and used to analyze low-frequency components, a<1 means that the wavelet is compressed and used to analyze high-frequency components, and b is the translation factor, which indicates the position of the wavelet on the time axis.

[0077] The feature extraction of high-frequency components specifically includes:

[0078] Extract time domain features of high-frequency components, including mean, standard deviation, crest factor, and kurtosis.

[0079] Use K-means clustering to identify abnormal events.

[0080] The mean describes the average amplitude of the signal and reflects the overall energy level of the signal.

[0081] The standard deviation measures the volatility of the signal. The larger the standard deviation, the more drastic the signal amplitude changes.

[0082] The crest factor is used to detect shock signals, and a higher crest factor may indicate the shock nature of the fault.

[0083] A larger kurtosis indicates that the signal has a stronger peak characteristic, which may be related to phenomena such as faults and arcs.

[0084] K-means clustering is used to classify the time domain features of high-frequency signals to identify possible abnormal events.

[0085] Set the number of categories, K, for K-means clustering and randomly initialize K cluster centers. Calculate the Euclidean distance from each sample to each cluster center, assign each sample to the nearest cluster center, recalculate the cluster center, take the average of all samples in the current category, and repeat the iteration until the cluster centers converge.

[0086] Calculate the distance distribution from the center point of each category to all sample points. If some data points are particularly far from the cluster center (i.e., outliers), they are considered abnormal.

[0087] The start and end time nodes of the abnormal event are extracted and recorded as the abnormal period. The real-time data stream within the abnormal period is intercepted and recorded as the abnormal data stream.

[0088] Modal decomposition of the abnormal data stream is performed to obtain several intrinsic mode functions, which are clustered to obtain high-frequency power data, medium-frequency power data and low-frequency power data, which are recorded as various modal power data.

[0089] It should be noted that an intrinsic mode function (IMF) is a signal component that meets specific conditions. Its components represent the frequency components of the original signal, arranged from high to low frequency. It characterizes the inherent vibrational patterns of the data and reflects the local characteristics of the signal at different time scales.

[0090] The empirical mode decomposition (EMD) method is used to decompose the abnormal data stream into several intrinsic mode functions (IMFs), which represent different frequency components. Specifically,

[0091] Find the extreme points of the abnormal data stream and identify all local maxima and minima. Fit the maximum and minimum points using cubic spline interpolation to form the upper and lower envelopes, respectively. Calculate the mean curve, which is the mean of the upper and lower envelopes. Calculate the signal minus the mean curve as the candidate IMF. If the candidate IMF meets the definition of an intrinsic mode function (i.e., locally symmetric and with a mean close to 0), record it as IMF1. Otherwise, continue iterating the candidate IMF until the conditions are met. Remove the extracted IMF and repeat the above steps until the remaining signal becomes a monotonic function or has low energy.

[0092] The decomposed IMF is subjected to feature extraction, and the main features include: center frequency, power spectrum density, instantaneous frequency and energy ratio.

[0093] The center frequency is the main frequency component for calculating the IMF and is used to distinguish high-frequency, medium-frequency, and low-frequency signals.

[0094] The power spectrum density is used to analyze the energy distribution of each IMF in different frequency bands and determine in which frequency range its energy is mainly concentrated.

[0095] The instantaneous frequency is used to calculate the instantaneous frequency characteristics of the IMF and analyze its dynamic changes.

[0096] The energy ratio is used to calculate the energy ratio of each IMF relative to the overall signal and determine its contribution to the signal.

[0097] Based on the extracted features, IMF is classified to determine which are high-frequency, medium-frequency and low-frequency power data. Specifically, they include:

[0098] The IMF frequency feature is selected as the clustering input, the number of cluster categories is set (3 in the embodiment of the present invention, corresponding to high frequency, medium frequency and low frequency), and the K-Means algorithm is run to divide the IMF into three categories.

[0099] After clustering is completed, the power data modes are divided according to the IMF categories:

[0100] High-frequency power data contains short-term sudden change signals, such as instantaneous harmonics and arc fault signals.

[0101] Medium frequency power data contains fault characteristic signals, such as partial discharge signals and short circuit signals.

[0102] Low-frequency power data represents the steady-state trends of the power grid, such as power fluctuations and voltage fluctuations.

[0103] The quality of each modal power data is detected to obtain quality detection data of each modal power data, including the noise power spectrum density, the amount of missing data and the amount of abnormal value data of each modal power data.

[0104] It should be noted that the definition and function of the quality detection data of each modal power data

[0105] In power data analysis, noise power spectral density (PSD) represents the noise power distribution of different frequency components in a signal. The unit is usually dB / Hz. Noise power spectral density is used to measure the noise level of the signal.

[0106] Missing data refers to the percentage of power data that was not collected during the sampling interval, typically expressed as the number of missing data points or the missing data percentage. It is used to assess data integrity. Missing data often affects analysis accuracy, and a high missing data percentage indicates lower data quality.

[0107] The amount of outlier data refers to the number of data points in power data that exceed the normal range or do not conform to physical laws.

[0108] The quality detection data of each modal power data are processed by averaging to obtain the mean value of the quality detection data, including the mean value of the noise power spectrum density, the mean value of the amount of missing data, and the mean value of the amount of outlier data. Based on the deviation value between the quality detection data of each modal power data and the mean value of the quality detection data, the quality detection value of each modal power data is obtained, specifically including:

[0109]

[0110] Among them, zl i is the quality detection value of the ith modal power data, PSD i is the noise power spectral density of the ith modal power data, MDA i is the amount of missing data for the ith modal power data, ODA i is the amount of abnormal value data of the ith modal power data, is the mean noise power spectral density, is the mean of the amount of missing data, is the mean value of the outlier data, α1 is the noise power spectrum density weight factor, α2 is the missing data weight factor, α3 is the outlier data weight factor, i is the modal number of the power data, i = 1, 2, 3, respectively, high-frequency power data, medium-frequency power data and low-frequency power data.

[0111] It should be noted that the noise power spectrum density weight factor, the missing data amount weight factor, and the outlier data amount weight factor all have values ranging from 0 to 1 and satisfy α1+α2+α3=1. The noise power spectrum density weight factor is an influence factor of the noise power spectrum density pre-stored in the built-in database of the data processing center, indicating the degree of influence of the noise power spectrum density on the quality detection value of each modal power data; the missing data amount weight factor is an influence factor of the missing data amount pre-stored in the built-in database of the data processing center, indicating the degree of influence of the missing data amount on the quality detection value of each modal power data; the outlier data amount weight factor is an influence factor of the outlier data amount pre-stored in the built-in database of the data processing center, indicating the degree of influence of the outlier data amount on the quality detection value of each modal power data. When used, they are directly extracted from the built-in database of the data processing center. For example, the noise power spectrum density, missing data amount, and outlier data amount of each modal power data are input into a preset mapping set in the built-in database of the data processing center to obtain the noise power spectrum density weight factor, missing data amount weight factor, and outlier data amount weight factor of each modal power data, and their corresponding mapping relationships are one-to-one corresponding.

[0112] It's also important to note that during power data quality monitoring, the noise power spectral density (PSD), the amount of missing data, and the amount of outlier data are interrelated. High noise levels typically increase the amount of outlier data. In high-noise environments, voltage and current signals often exhibit sudden or abnormal fluctuations, leading to an increase in outlier data. For example, high-frequency noise can cause significant offsets in voltage or current signals at certain moments, leading the outlier detection algorithm to identify them as anomalous data. Low-frequency noise often causes overall signal drift, causing the data to deviate from the normal range. An increase in outlier data typically indicates high noise power. A high proportion of outliers in the data typically indicates strong signal interference or decreased measurement equipment accuracy. This is often reflected by an abnormal increase in energy in specific frequency bands on the power spectral density plot, such as an increase in specific harmonic components caused by electromagnetic interference. Short-term high-frequency sudden changes can lead to an abnormal increase in high-frequency PSD. High noise levels often result in missing data. In high-noise environments, sensors or data acquisition equipment often fail to correctly identify valid signals, resulting in data loss. For example, interference in the communication link can cause data packet loss and intermittent data loss. Sensor signals are covered by noise, making it impossible to interpret valid data and recording them as null values (NaNs). Increased amounts of missing data often affect PSD calculations. A large amount of missing data can lead to energy leakage or spectral distortion when calculating the PSD using the Fourier transform (FFT), resulting in inaccurate noise estimation. For example, missing data can cause discontinuities in the time domain signal, resulting in the appearance of spurious high-frequency components in the FFT calculation. Missing data requires interpolation, which often introduces new errors and distorts the PSD calculation. Excessive outliers often result in missing data. Many data processing systems, upon detecting outliers, delete or fill them with "NaNs," leading to an increase in the amount of missing data. If the frequency of outliers is too high within a short period of time, the device often enters protection mode and stops recording data, resulting in missing data. Excessive amounts of missing data often hinder outlier detection. Because outlier detection methods (such as the 3σ rule and isolation forest) rely on data distribution characteristics, excessive amounts of missing data often lead to inaccurate anomaly detection. Statistical methods (such as the mean and standard deviation) often deviate from the true value when data is missing, thus affecting the accuracy of outlier determination.

[0113] Based on the quality detection value of each modal power data, a mapping match is performed with the proportional constraint factor corresponding to the interval of each quality detection value pre-stored in the data processing center to obtain the proportional constraint factor of each modal power data and perform the first proportional adjustment on each modal power data.

[0114] S2. Send the abnormal data stream to the data processing center to obtain the fault node, extract the modal power data of each adjacent node of the fault node for analysis, and determine whether it causes a diffuse fault.

[0115] The abnormal data stream is sent to the data processing center for analysis to obtain the fault node. The specific process is as follows:

[0116] The abnormal data stream is sent to the data processing center, which extracts abnormal current data and abnormal voltage data and determines the abnormal type, which includes single-phase grounding fault, double-phase grounding fault, double-phase short circuit fault and three-phase short circuit fault.

[0117] The specific process of determining the abnormality type includes:

[0118] Based on the abnormal current data and abnormal voltage data, the phase current and phase voltage are calculated to obtain the amplitude and phase changes of the three-phase voltage and current.

[0119] The specific steps are:

[0120] Step 1: Calculate the zero-sequence current and negative-sequence current to determine whether an asymmetric fault exists.

[0121] Step 2: Determine whether the current or voltage of a phase is significantly reduced or approaches zero.

[0122] Step 3: Determine whether two or three phases have abnormal changes at the same time.

[0123] Step 4: Observe the magnitude of the zero-sequence component and the negative-sequence component to determine the fault type.

[0124] Step 5: Determine the fault type based on the current and voltage change patterns and compare them with standard fault characteristics:

[0125] The standard fault characteristics specifically include:

[0126] The main characteristics of a single-phase ground fault are a decrease in or near-zero voltage on one phase, a significant increase in zero-sequence current, and an increase in zero-sequence voltage. The voltages of the other two phases are relatively high. When these characteristics are detected, it can be identified as a single-phase ground fault.

[0127] The main characteristics of a double-phase ground fault are a significant drop in voltage or near-zero on both phases, a significant increase in both zero-sequence current and voltage, and an increase in the voltage on the other phase. When these characteristics are detected, it can be identified as a double-phase ground fault.

[0128] The main characteristics of a two-phase short-circuit fault are a sharp increase in both-phase currents, a significant rise in the short-circuit current, a significant increase in the negative-sequence component, and a small or absent zero-sequence component. The voltages on the two short-circuited phases decrease significantly or approach zero. When these characteristics are detected, a two-phase short-circuit fault can be identified.

[0129] The main characteristics of a three-phase short-circuit fault are a simultaneous sharp increase in all three-phase currents and a significant decrease or near-zero decrease in all three-phase voltages. The short-circuit current is extremely high. When these characteristics are detected, a three-phase short-circuit fault can be identified.

[0130] Based on the abnormal current data and abnormal voltage data, combined with the abnormal type, the fault node is obtained by impedance analysis.

[0131] The impedance method studies circuit characteristics by measuring impedance. Depending on the application, impedance (Z) is a complex number representing the circuit's resistance to AC current. It consists of a real part (resistance R) and an imaginary part (reactance X). Reactance is further divided into inductive reactance (the impedance of an inductor) and capacitive reactance (the impedance of a capacitor). The impedance method does not require destructive processing and can detect even small impedance changes, making it suitable for high-precision measurements.

[0132] Obtaining the faulty node specifically includes:

[0133] Based on the real-time voltage and real-time current of each power node, the impedance of each power node is obtained by ratio processing, and then compared with the rated impedance of each power node pre-stored in the built-in database of the data processing center to obtain the impedance deviation value. If the impedance deviation value of a power node is greater than or equal to the impedance deviation threshold, the node is determined to be a faulty node.

[0134] It should be noted that the characteristic impedance of each line segment, ie, the rated impedance, is stored in the built-in database of the data processing center.

[0135] The real-time data streams of the adjacent nodes of the faulty node are extracted and analyzed to obtain the fault diagnosis results of the adjacent nodes, which specifically include:

[0136] The voltage, current, frequency, active power and harmonic distortion rate in the real-time data stream of each adjacent node of the faulty node are extracted, and the voltage, current, frequency, active power and harmonic distortion rate in the normal data stream of each adjacent node of the faulty node are retrieved from the built-in database of the data processing center. The comparison is performed to obtain the diagnostic deviation value. The fault diagnosis results of each adjacent node of the faulty node are obtained based on the analysis and processing of the diagnostic deviation value, which specifically include:

[0137]

[0138] ΔV h =|V h,s -V h,before |;

[0139] ΔI h =|I h,s -I h,before |;

[0140] Δfh =|f h,s -f h,before |;

[0141] ΔP h =|P h,s -P h,before |;

[0142] ΔTHD h =|THD h,s -THD h,before |;

[0143] Among them, JG h is the diagnostic deviation value of the hth adjacent node of the faulty node, ΔV h is the voltage diagnosis deviation value of the hth adjacent node of the fault node, ΔI h is the current diagnosis deviation value of the hth adjacent node of the fault node, Δf h is the frequency diagnosis deviation value of the hth adjacent node of the faulty node, ΔP h is the active power diagnosis deviation value of the hth adjacent node of the faulty node, ΔTHD h is the harmonic distortion rate diagnostic deviation value of the hth adjacent node of the faulty node, V0 is the voltage in the normal data stream, I0 is the current in the normal data stream, f0 is the frequency in the normal data stream, P0 is the active power in the normal data stream, THD0 is the harmonic distortion rate in the normal data stream, V h,s is the real-time voltage of the hth adjacent node of the fault node, I h,s is the real-time current of the hth adjacent node of the fault node, f h,s is the real-time frequency of the hth neighboring node of the faulty node, P h,s is the real-time active power of the hth neighboring node of the faulty node, THD h,s is the real-time harmonic distortion rate of the hth adjacent node of the faulty node, V h,before is the voltage of the hth adjacent node of the fault node before the abnormal period, I h,before is the current of the hth adjacent node of the fault node before the abnormal period, f h,before is the frequency of the hth neighboring node of the faulty node before the abnormal period, P h,before is the active power of the hth adjacent node of the faulty node before the abnormal period, THD h,before is the harmonic distortion rate of the hth adjacent node of the faulty node before the abnormal period, h is the adjacent node number of the faulty node, h = 1, 2, 3, ..., n, n is the total number of adjacent nodes of the faulty node, δ1 is the voltage weight, δ2 is the current weight, δ3 is the frequency weight, δ4 is the active power weight, and δ5 is the harmonic distortion rate weight.

[0144] It should be noted that the voltage weight, current weight, frequency weight, active power weight and harmonic distortion rate weight all have value ranges between 0 and 1, and satisfy δ1+δ2+δ3+δ4+δ5=1. The voltage weight is the influence factor of the voltage pre-stored in the built-in database of the data processing center, which indicates the degree of influence of the voltage on the diagnostic deviation value of each adjacent node of the fault node; the current weight is the influence factor of the current pre-stored in the built-in database of the data processing center, which indicates the degree of influence of the current on the diagnostic deviation value of each adjacent node of the fault node; the frequency weight is the influence factor of the frequency pre-stored in the built-in database of the data processing center, which indicates the degree of influence of the frequency on the diagnostic deviation value of each adjacent node of the fault node; The active power weight is an influencing factor of active power pre-stored in the built-in database of the data processing center, which indicates the degree of influence of active power on the diagnostic deviation value of each adjacent node of the faulty node; the harmonic distortion rate weight is an influencing factor of harmonic distortion rate pre-stored in the built-in database of the data processing center, which indicates the degree of influence of harmonic distortion rate on the diagnostic deviation value of each adjacent node of the faulty node. When used, it is directly extracted from the built-in database of the data processing center. For example, the node voltage, current, frequency, active power and harmonic distortion rate are input into the preset mapping set in the built-in database of the data processing center to obtain the voltage weight, current weight, frequency weight, active power weight and harmonic distortion rate weight, and the corresponding mapping relationship is one-to-one.

[0145] It's important to note that there's a correlation between voltage, current, frequency, active power, and harmonic distortion in data streams. Under constant load, a voltage decrease results in a decrease in active power. For example, when grid voltage decreases (e.g., due to increased load or transformer voltage drop), power supply capacity decreases, impacting the normal operation of load equipment. Conversely, a voltage increase typically leads to grid overload, shortening the lifespan of grid equipment. Under constant voltage conditions, an increase in current increases active power, but also increases line losses. When load increases, the system's total active power demand rises. If generator output is insufficient, the frequency decreases. Conversely, when load decreases and power generation is excessive, the frequency increases. Harmonic currents flowing through grid impedances cause voltage harmonic distortion, impacting power quality. Harmonic components typically reduce the actual transmitted active power, affecting the effective power utilization of the load. Voltage and current determine active power and also affect equipment operating efficiency and losses. Frequency reflects the power balance of the power system. When load increases, the frequency decreases, often impacting equipment operation. THD reflects the quality of electric energy. Excessive harmonics will affect the voltage and current waveforms, causing damage to power equipment or reducing the power factor.

[0146] The fault diagnosis results of the adjacent nodes of the faulty node include node fault and node normal.

[0147] When the diagnosis deviation value of a certain adjacent node is greater than or equal to the diagnosis deviation threshold, the fault diagnosis result of the adjacent node is determined to be a node fault.

[0148] When the diagnosis deviation value of a certain adjacent node is less than the diagnosis deviation threshold, the fault diagnosis result of the adjacent node is determined to be normal.

[0149] If the fault diagnosis result of an adjacent node of the faulty node is a node fault, it is determined that the faulty node causes a diffuse fault.

[0150] The faulty node and the adjacent nodes diagnosed as node faults are collectively referred to as faulty nodes.

[0151] If the fault diagnosis result of an adjacent node of the faulty node is that the node is normal, it is determined that the faulty node has not caused a diffuse fault and is recorded as a single faulty node.

[0152] The adjacent nodes whose fault diagnosis results are node failures are recorded as each faulty second-level node, and the real-time data streams of the other adjacent nodes of each faulty second-level node are retrieved for fault diagnosis. If there are other adjacent nodes with the fault diagnosis results of node failures, they are recorded as each faulty third-level node, and so on, until the fault diagnosis results of the other adjacent nodes of each faulty N-level node are all nodes that are normal, and the faulty node and each faulty N-level node are counted as each faulty node.

[0153] It should be noted that the other adjacent nodes refer to nodes adjacent to the faulty node except the nodes that have been diagnosed as faulty.

[0154] S3. If a diffusion fault occurs, the hardware performance of the built-in data processing module of each fault node is tested based on the data volume of each modal power data of each fault node. If the hardware performance test result is qualified, the modal power data after adjusting the proportional constraint factor is input into the local processor under the built-in data processing module of each fault node for analysis to obtain the first monitoring adjustment parameter of each fault node for adjustment.

[0155] Based on the amount of power data of each mode at each fault node, the hardware performance test of the built-in data processing module of each fault node is performed, specifically including:

[0156] The power data of each mode of each fault node during the abnormal period is retrieved, and after local pre-processing, statistics are performed to obtain the data volume of the power data of each mode of each fault node.

[0157] Obtain hardware performance test data of the built-in data processing module of each faulty node, including the number of CPU threads, CPU instruction execution rate, memory and HDD read and write rates of the built-in data processing module of each faulty node.

[0158] It's important to note that the number of CPU threads refers to the number of threads executing tasks simultaneously within the processor, that is, the number of tasks the CPU can handle in parallel. The greater the number of threads, the better the CPU's performance in multitasking and parallel computing. In power data analysis, a high number of CPU threads allows for faster parallel processing of multiple data streams, accelerating fault analysis.

[0159] The CPU instruction execution rate, typically expressed as instructions per second (IPS) or floating-point operations per second (FLOPS), reflects the speed at which the CPU performs computational tasks. The higher the instruction execution rate, the more computational tasks the CPU can complete per unit time. In model application scenarios, a higher instruction execution rate can accelerate data preprocessing, feature extraction, model calculation, and other processes. However, a low CPU instruction execution rate often results in computational delays in the data processing module under high load, impacting fault identification and recovery.

[0160] Memory (RAM, Random Access Memory) is a component used by the CPU to store temporary data when running a program, which determines the storage capacity and access speed during data processing. The size of memory is usually measured in GB (gigabytes) or TB (terabytes). During the power data analysis process, the data needs to be stored in the memory before the CPU can process it. If the memory capacity is insufficient, the system will usually use disk storage frequently, resulting in slower calculations. Power grid data analysis requires processing large amounts of high-dimensional data, and the inference and training of the model have high memory requirements. A larger memory capacity can avoid overflow and improve computing stability. A larger RAM can store data streams for longer periods of time, ensuring that the system can efficiently handle abnormal events without losing key information due to insufficient cache.

[0161] Disk read / write speed refers to the speed at which data is read from or written to a hard disk drive (HDD), typically measured in MB / s (megabytes per second) or IOPS (input / output operations per second). Power grid fault data typically needs to be stored in a database and quickly accessed when needed. Low disk read / write speeds often result in slow data loading, impacting analysis efficiency. In model calculations, model training and inference typically require reading large amounts of data from disk. Higher read / write speeds can reduce I / O bottlenecks and improve overall system performance.

[0162] Based on the data volume of each modal power data of each fault node, a mapping and matching is performed with each hardware performance detection verification set corresponding to the data volume interval of each modal power data pre-stored in the built-in database of the data processing center to obtain the hardware performance detection verification set of each fault node, including the number of CPU thread verification, CPU instruction execution verification rate, verification memory and HDD read and write verification rate.

[0163] The hardware performance test data of the built-in data processing module of each faulty node is compared and analyzed with the hardware performance test verification set of each faulty node to obtain the hardware performance test results of the built-in data processing module of each faulty node, specifically including:

[0164]

[0165] Among them, XN y is the hardware performance monitoring value of the built-in data processing module of the yth fault node, CPUTC y is the number of CPU threads in the built-in data processing module of the yth fault node, CPUIER y is the CPU instruction execution rate of the built-in data processing module of the yth fault node, RAM y is the memory of the built-in data processing module of the yth fault node, HDDWS y is the HDD read and write rate of the built-in data processing module of the yth fault node, CPUTC y,0 is the number of CPU thread verifications of the built-in data processing module of the yth fault node, CPUIER y,0 The CPU instruction execution check rate of the built-in data processing module of the yth fault node, RAM y,0 The checksum memory of the built-in data processing module of the yth fault node, HDDWS y,0 is the HDD read / write verification rate of the built-in data processing module of the yth faulty node, γ1 is the weight factor of the number of CPU threads, γ2 is the weight factor of the CPU instruction execution rate, γ3 is the memory weight factor, γ4 is the HDD read / write rate weight factor, y is the faulty node number, y = 1, 2, 3, ..., m, where m is the total number of faulty nodes.

[0166] It should be noted that the CPU thread quantity weight factor, CPU instruction execution rate weight factor, memory weight factor and HDD read / write rate weight factor all have value ranges between 0 and 1, and satisfy γ1+γ2+γ3+γ4=1. The CPU thread quantity weight factor is the influence factor of the CPU thread quantity pre-stored in the built-in database of the data processing center, which indicates the degree of influence of the CPU thread quantity on the hardware performance monitoring value of the built-in data processing module of each fault node; the CPU instruction execution rate weight factor is the influence factor of the CPU instruction execution rate pre-stored in the built-in database of the data processing center, which indicates the degree of influence of the CPU instruction execution rate on the hardware performance monitoring value of the built-in data processing module of each fault node; the memory weight factor is the influence factor of electromagnetic interference pre-stored in the built-in database of the data processing center. The sub-factor represents the degree of influence of the memory on the hardware performance monitoring value of the built-in data processing module of each faulty node; the HDD read / write rate weight factor is an influence factor of the HDD read / write rate pre-stored in the built-in database of the data processing center, which represents the degree of influence of the HDD read / write rate on the hardware performance monitoring value of the built-in data processing module of each faulty node. When used, it is directly extracted from the built-in database of the data processing center. For example, the real-time CPU thread number, real-time CPU instruction execution rate, real-time electromagnetic interference and real-time HDD read / write rate of the built-in data processing module of each faulty node are input into the preset mapping set in the built-in database of the data processing center to obtain the CPU thread number weight factor, CPU instruction execution rate weight factor, electromagnetic interference weight factor and HDD read / write rate weight factor, and the corresponding mapping relationship is one-to-one correspondence.

[0167] It should also be noted that the number of CPU threads, CPU instruction execution rate, and memory and HDD read / write rates of the built-in data processing module of each fault node are closely related. Multithreading can improve parallel computing capabilities. When multiple threads run simultaneously, the CPU can handle more tasks. The instruction execution rate determines the computing efficiency of a single thread. Even with many threads, a low instruction execution rate will still limit the overall computing speed. For power data processing, computing tasks may involve signal analysis, machine learning inference, and other tasks. A high instruction execution rate and multithreading support can improve computing throughput and reduce data processing latency. CPU computing relies on memory for data exchange. If the memory is too small or the access speed is slow, even a powerful CPU may suffer from reduced computing efficiency due to data latency (cache misses). Data processing tasks often require large-scale memory access. For example, power grid fault detection involves the storage and calculation of large-scale time-series data. If memory is insufficient, the system may rely on HDDs (or SSDs) for virtual memory exchange, reducing overall processing efficiency. If the memory is large enough, the CPU can cache more data, reducing the number of HDD or SSD accesses and improving data processing efficiency. When memory is insufficient, the system uses the HDD for data exchange (virtual memory). Slow HDD read and write speeds can create data exchange bottlenecks, impacting overall performance. HDD access speed affects data loading time. For example, when the system needs to read large amounts of historical electricity data, faster HDD read and write speeds maximize CPU and memory computing efficiency. Using an SSD (solid-state drive) significantly increases data read speeds. Compared to traditional HDDs, SSDs offer faster random read and write speeds, reducing CPU wait time and improving data processing efficiency.

[0168] If the hardware performance test result of the built-in data processing module of a faulty node is qualified, the power data of each mode of the faulty node will be analyzed and processed.

[0169] If the hardware performance test result of the built-in data processing module of a faulty node is unqualified, the demand processing data of the faulty node is imported into the data processing center for analysis and processing.

[0170] If the hardware performance test result is qualified, the modal power data after adjusting the proportional constraint factor is input into the local processor under the built-in data processing module of each fault node for analysis to obtain the first monitoring adjustment parameter of each fault node and adjust the monitoring period and data extraction frequency of each fault node, specifically including:

[0171] The power data of each mode of each fault node is imported into the local processor for processing and analysis. The specific processing process includes:

[0172] Based on the proportional constraint factor after the first adjustment, the local processor adjusts the adoption ratio of each modal power data, and extracts abnormal current data and abnormal voltage data therefrom, determines the abnormal type of each fault node, and based on the fault type of each fault node, matches it with the fault factors corresponding to each fault type pre-stored in the local processor to obtain the fault factor of each fault node, which is used to characterize the severity of the fault type.

[0173] Based on the fault factor of each fault node, a first monitoring adjustment parameter corresponding to the fault factor is extracted, and the first monitoring adjustment parameter of each fault node is obtained to adjust the monitoring period and data extraction frequency of each fault node.

[0174] The first monitoring adjustment parameter of each fault node includes a monitoring period length adjustment parameter and a data extraction frequency adjustment parameter.

[0175] S4. If no diffusion fault is caused, the faulty device is identified, the data mode to which the faulty device belongs is obtained, the number of faulty devices in each mode of power data is counted, the proportional constraint factor of each mode of power data is adjusted for a second time, and the data is imported into the data processing center for analysis and processing to obtain the second monitoring adjustment parameter of the single fault node and adjust the monitoring period of the single fault node.

[0176] If no diffusion fault occurs, the faulty device is identified, the data mode to which the faulty device belongs is obtained, the number of faulty devices in each mode of power data is counted, and the proportional constraint factor of each mode of power data is adjusted for a second time, specifically including:

[0177] Collect statistics on the real-time performance status parameters and historical usage records of each device at a single fault node.

[0178] The real-time performance status parameters include real-time load current, real-time vibration signal and harmonic content.

[0179] The historical usage records include the cumulative time of use, number of failures and number of types of failures.

[0180] It should be noted that the real-time load current of the transformer refers to the instantaneous current value passing through its winding during the operation of the transformer, which is usually measured in amperes (A). The larger the load current, the higher the working pressure of the transformer.

[0181] Real-time vibration signal refers to the mechanical vibration amplitude generated during the operation of transformers or other power grid equipment, usually expressed in acceleration (m / s 2 ) and is monitored in real time by vibration sensors.

[0182] Harmonics refer to current or voltage components in the power grid that have frequencies other than the fundamental frequency (50Hz or higher). They are typically measured by total harmonic distortion (THD), expressed as a percentage. High harmonic content can degrade power quality and affect the proper operation and lifespan of equipment.

[0183] Real-time performance status verification parameters and historical usage record verification parameters are extracted from the built-in database of the data processing center.

[0184] The real-time performance status verification parameters include the transformer real-time load verification current, the real-time verification vibration signal and the verification harmonic content.

[0185] The historical usage record verification parameters include a verification factor for the cumulative length of time in use, a number of fault verifications, and a number of fault type verifications.

[0186] Based on the real-time performance status parameters and historical usage records of each device at a single fault node, the load index of each device is obtained through analysis and processing, including:

[0187]

[0188] Among them, δ r is the load index of the rth device, I r is the real-time load current of the rth device, vs r is the real-time vibration signal of the rth device, thd r is the harmonic content of the rth device, Tm r gT is the cumulative time of the rth device in use. r is the number of failures of the rth device, q r is the number of fault types of the rth device, I0 is the real-time load verification current of the transformer, vs0 is the real-time verification vibration signal, thd0 is the verification harmonic content, ε t is the verification factor for the cumulative time of use, gT0 is the number of fault verifications, q0 is the number of fault type verifications, ω1 is the real-time performance status weight factor, ω2 is the historical usage record weight factor, r is the number of devices in a single fault node, r = 1, 2, 3, ..., L, and L is the total number of devices in a single fault node.

[0189] It should be noted that the real-time performance status weight factor and the historical usage record weight factor both have a value range between 0 and 1, and satisfy ω1+ω2=1. The real-time performance status weight factor is an influencing factor of the real-time performance status pre-stored in the built-in database of the data processing center, indicating the degree of influence of the real-time performance status on the load index of each device of a single fault node; the historical usage record weight factor is an influencing factor of the historical usage record pre-stored in the built-in database of the data processing center, indicating the degree of influence of the historical usage record on the load index of each device of a single fault node. When used, it is directly extracted from the built-in database of the data processing center. For example, the real-time performance status and historical usage records of each device of a single fault node are input into the preset mapping set in the built-in database of the data processing center to obtain the real-time performance status weight factor and historical usage record weight factor of each device of a single fault node, and the corresponding mapping relationship is one-to-one.

[0190] It should also be noted that the real-time load current, real-time vibration signal, harmonic content, cumulative service time, number of faults and number of fault types are related when evaluating the sinusoidal excitation current carrying index of the equipment. The larger the load current, the higher the electromagnetic stress of the equipment, and the internal winding may be subjected to a greater thermal load, which in turn affects the carrying capacity of the equipment.

[0191] Fluctuations in load current may cause additional stress on the device's internal structure, affecting its response to sinusoidal excitation current. When evaluating the sinusoidal excitation current carrying capacity index, if the device is under high load, the test current amplitude may need to be reduced to avoid overloading the device. An abnormal increase in vibration signals may indicate device aging or damage, potentially reducing its ability to withstand additional sinusoidal excitation current. During device operation, if the vibration signal exceeds the safety threshold, the sinusoidal excitation current amplitude may need to be reduced to prevent further damage. Historical vibration signal data can be used to determine whether the device has experienced long-term mechanical vibration issues and adjust the sinusoidal excitation current testing strategy accordingly. High harmonic content can cause device heating, insulation aging, and additional electromagnetic interference, which in turn affect the device's ability to withstand sinusoidal excitation current. In the presence of high harmonic content, applying additional sinusoidal excitation current may cause the device to enter a nonlinear operating state, affecting the accuracy of test results. If the device is exposed to a high-harmonic environment for a long time, the sinusoidal excitation current amplitude may need to be reduced to prevent a decrease in the device's carrying capacity. Equipment that has been in operation for a long time often experiences material aging and insulation degradation, and may require a lower sinusoidal excitation current to avoid additional stress. New equipment typically has a stronger load-carrying capacity and can withstand larger sinusoidal excitation currents for testing. Combined with usage duration data, test strategies can be optimized. Equipment that frequently fails may have structural issues, lower load-carrying capacity, and may be less tolerant to sinusoidal excitation currents. When testing equipment with a high number of failures, it may be necessary to reduce the test current or select a different excitation frequency to reduce additional losses and impact. If a device has previously failed under certain load conditions, the test strategy can be adjusted based on historical data to prevent similar situations from recurring. Equipment that experiences multiple types of failures may have more complex damage mechanisms, requiring greater caution when assessing its sinusoidal excitation current carrying capacity. For example, if a device has previously experienced insulation damage, special attention should be paid to its voltage withstand capability when performing a sinusoidal excitation current test to avoid triggering the fault again. Combined with fault type data, different excitation waveforms can be selected (for example, reducing the high-frequency component) or the excitation amplitude can be reduced to improve test safety.

[0192] And based on the load index of each device and the load index in the built-in database of the data processing center, the sinusoidal excitation current size corresponding to each device is mapped and matched to obtain the sinusoidal excitation current size of each device.

[0193] A sinusoidal excitation current is generated for each device, and the faulty device is identified based on the response data of each device. Specifically, the actual response data of the device is compared with the response data of a standard healthy device. If there are abnormal changes (such as increased harmonics, current distortion, and abnormal power factor), the device is identified as faulty.

[0194] It should be noted that the sinusoidal excitation current is a current applied to an electrical device with a sinusoidal waveform of a specific frequency, amplitude, and phase. This current is typically used to test the dynamic response characteristics of electrical equipment to identify its health status or fault conditions.

[0195] By applying a sinusoidal excitation current to power equipment (such as transformers, motors, circuit breakers, etc.), the response of the equipment to currents of different frequencies can be detected, thereby analyzing the health status of the equipment.

[0196] During normal operation, the equipment's response to the excitation current should conform to expected characteristics. If an anomaly (such as winding aging, short circuits, or poor contact) occurs, the response curve will be distorted. For example, nonlinear distortion in the motor's response current indicates a stator winding short circuit, while increased harmonic content in the transformer suggests core saturation or insulation degradation.

[0197] Perform modal separation and feature extraction on the output data of each faulty device to obtain the main output mode of each faulty device, and count the number of faulty devices in each modal power data.

[0198] Based on the ratio of the number of faulty devices in each modal power data to the total number of faulty devices, a second adjustment is made to the proportional constraint factor of each modal power data in the abnormal period.

[0199] The data is imported into the data processing center for analysis and processing to obtain the second monitoring adjustment parameter of the single fault node and adjust the monitoring period of the single fault node. The specific processing conditions are:

[0200] Each modal power data is imported into the data processing center. Based on the secondary adjustment ratio constraint factor, the adoption ratio of each modal power data is adjusted. After feature extraction of each modal power data, the fault weight of each feature is extracted from the built-in database of the data processing center. The fault weight is used to characterize the fault severity of the feature and comprehensively evaluate the severity of a single fault node, specifically including:

[0201]

[0202] Among them, G is the fault severity assessment value of a single fault node, Nm u is the failure factor of the u-th feature, is the fault weight factor of the u-th feature, u is the feature number of the single fault node, u=1,2,3,...,sy, sy is the total number of features of the single fault node.

[0203] Based on the fault severity assessment value of a single fault node, a mapping and matching is performed with the monitoring period adjustment parameters corresponding to the intervals of each fault severity assessment value pre-stored in the built-in database of the data processing center to obtain the monitoring period adjustment parameters of the single fault node and focus on monitoring the single fault node.

[0204] The monitoring period adjustment parameters of the single fault node include a monitoring period length adjustment parameter and a data extraction frequency adjustment parameter.

[0205] In this embodiment, the present invention provides a multimodal power data analysis system based on an AI large model, comprising:

[0206] The high-frequency modal monitoring module is used to input the real-time data stream in the power grid into the AI large model, monitor the high-frequency modes in the power grid in real time and identify abnormal events, and make the first adjustment to the proportional constraint factor of each modal power data based on the data quality of each modal power data in the abnormal data stream.

[0207] The diffusion fault judgment module is used to send the abnormal data stream to the data processing center to obtain the fault node, extract the modal power data of each adjacent node of the fault node for analysis, obtain the fault diagnosis results of each adjacent node and determine whether a diffusion fault is caused.

[0208] The multi-node fault processing module is used to perform hardware performance testing on the built-in data processing module of each fault node based on the data volume of each modal power data of each fault node if a diffusion fault occurs. If the hardware performance test result is qualified, the modal power data after adjusting the proportional constraint factor is input into the local processor under the built-in data processing module of each fault node for analysis to obtain the first monitoring adjustment parameter of each fault node and adjust the monitoring period and data extraction frequency of each fault node.

[0209] The single-node fault processing module is used to identify the faulty device if no diffusion fault occurs, obtain the data mode to which the faulty device belongs, count the number of faulty devices in each mode of power data, make a second adjustment to the proportional constraint factor of each mode of power data, and import it into the data processing center for analysis and processing to obtain the second monitoring adjustment parameter of the single fault node and adjust the monitoring period of the single fault node.

[0210] In addition, a multimodal power data analysis device based on an AI large model is provided, wherein the device has one or more programs, and the one or more programs are executed by one or more processors to implement the above method.

[0211] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0212] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made based on the contents of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention. As long as they do not deviate from the structure of the present invention or exceed the scope defined by the present invention, they should fall within the scope of protection of the present invention.

Claims

1. A multimodal power data analysis method based on AI large model, characterized in that: include: S1. Input the real-time data stream from the power grid into the AI large model, monitor the high-frequency modes in the power grid in real time and identify abnormal events. Based on the data quality of each modal power data in the abnormal data stream, the proportional constraint factor of each modal power data is adjusted for the first time; S2. Send the abnormal data stream to the data processing center to obtain the fault node, extract the modal power data of each adjacent node of the fault node for analysis, and determine whether it causes a diffuse fault; S3. If a diffusion fault occurs, a hardware performance test is performed on the built-in data processing module of each fault node based on the data volume of each modal power data of each fault node. If the hardware performance test result is qualified, the modal power data after adjusting the proportional constraint factor is input into the local processor of the built-in data processing module of each fault node for analysis to obtain the first monitoring adjustment parameter of each fault node for adjustment; S4. If no diffusion fault is caused, the faulty device is identified, the data mode to which the faulty device belongs is obtained, the number of faulty devices in each mode of power data is counted, the proportional constraint factor of each mode of power data is adjusted for a second time, and the data is imported into the data processing center for analysis and processing to obtain the second monitoring adjustment parameter of the single fault node and adjust the monitoring period of the single fault node.

2. The multimodal power data analysis method based on the AI large model according to claim 1 is characterized by: The first adjustment of the proportional constraint factor of each modal power data based on the data quality of each modal power data in the abnormal data stream is performed as follows: The real-time data stream of the power grid is obtained through the synchronized phasor measurement unit of the PMU, including the real-time voltage, real-time current, real-time phase and real-time frequency of the power grid; The real-time data stream of the power grid is input into the AI large model for time-frequency analysis to obtain the spectrum of the real-time data stream. The spectrum is modally decomposed to separate the high-frequency components, and features are extracted from the high-frequency components. Abnormal events are identified based on the extracted features. Extract the start and end time nodes of the abnormal event, record them as the abnormal period, and intercept the real-time data stream within the abnormal period, record them as the abnormal data stream; Perform modal decomposition on the abnormal data stream to obtain several intrinsic mode functions, cluster the intrinsic mode functions to obtain high-frequency power data, medium-frequency power data and low-frequency power data, which are recorded as various modal power data; Performing quality detection on each modal power data to obtain quality detection data of each modal power data, including the noise power spectrum density, the amount of missing data, and the amount of abnormal value data of each modal power data; performing mean processing on the quality detection data of each modal power data to obtain a quality detection data mean, including the mean of the noise power spectrum density, the mean of the amount of missing data, and the mean of the amount of abnormal value data; based on the deviation value of the quality detection data of each modal power data and the mean of the quality detection data, correcting the coupling to obtain the quality detection value of each modal power data; Based on the quality detection value of each modal power data, a mapping match is performed with the proportional constraint factor corresponding to the interval of each quality detection value pre-stored in the data processing center to obtain the proportional constraint factor of each modal power data and perform the first adjustment.

3. The multimodal power data analysis method based on the AI large model according to claim 1 is characterized by: The abnormal data stream is sent to the data processing center to obtain the fault node. The specific process is as follows: The abnormal data stream is sent to the data processing center, which extracts the abnormal current data and abnormal voltage data and determines the abnormal type; Based on abnormal current data and abnormal voltage data, combined with the abnormal type, the impedance method is used to analyze the fault node; Obtaining the faulty node specifically includes: Based on the real-time voltage and real-time current of each power node, the impedance of each power node is processed and compared with the rated impedance of each power node pre-stored in the built-in database of the data processing center to obtain the impedance deviation value. If the impedance deviation value of a power node is greater than or equal to the impedance deviation threshold, the node is determined to be a faulty node.

4. The multimodal power data analysis method based on the AI large model according to claim 1 is characterized by: The extracting and analyzing the modal power data of each adjacent node of the faulty node to determine whether a diffusion fault is caused specifically includes: The real-time data streams of the nodes adjacent to the faulty node are extracted, and the normal data streams of the nodes adjacent to the faulty node are retrieved from the built-in database of the data processing center. After comparison and coupling, the diagnostic deviation value is obtained. Based on the analysis and processing of the diagnostic deviation value, the fault diagnosis results of the nodes adjacent to the faulty node are obtained: The fault diagnosis results of the adjacent nodes of the faulty node include node fault and node normal; When the diagnostic deviation value of a certain adjacent node is greater than or equal to the diagnostic deviation threshold, the fault diagnosis result of the adjacent node is determined to be a node fault; When the diagnostic deviation value of a certain adjacent node is less than the diagnostic deviation threshold, the fault diagnosis result of the adjacent node is determined to be normal; If the fault diagnosis result of a neighboring node of the faulty node is a node fault, it is determined that the faulty node causes a diffuse fault; The faulty node and the adjacent nodes diagnosed as node faults are collectively referred to as faulty nodes; If the fault diagnosis result of an adjacent node of the faulty node is that the node is normal, it is determined that the faulty node has not caused a diffuse fault and is recorded as a single faulty node.

5. The multimodal power data analysis method based on the AI large model according to claim 1 is characterized by: If a diffusion fault occurs, the hardware performance test of the built-in data processing module of each fault node is performed based on the data volume of each modal power data of each fault node, specifically including: Retrieve the power data of each mode of each fault node during the abnormal period, perform local pre-processing and statistics to obtain the data volume of each mode of power data of each fault node; Obtain hardware performance test data of the built-in data processing module of each faulty node, including the number of CPU threads, CPU instruction execution rate, memory and HDD read and write rates of each faulty node's built-in data processing module; Based on the data volume of each modal power data of each faulty node, a mapping match is performed with the hardware performance test verification set corresponding to the data volume range of each modal power data pre-stored in the built-in database of the data processing center. The hardware performance test verification set of each faulty node is obtained, including the number of CPU thread verification, CPU instruction execution verification rate, and verification memory and HDD read and write verification rates; Compare and analyze the hardware performance test data of the built-in data processing module of each faulty node with the hardware performance test check set of each faulty node to obtain the hardware performance test results of the built-in data processing module of each faulty node; If the hardware performance test result of the built-in data processing module of a faulty node is qualified, the power data of each mode of the faulty node will be analyzed and processed; If the hardware performance test result of the built-in data processing module of a faulty node is unqualified, the demand processing data of the faulty node is imported into the data processing center for analysis and processing.

6. The multimodal power data analysis method based on the AI large model according to claim 1 is characterized by: The first monitoring adjustment parameter of each fault node obtained by the analysis is adjusted, specifically including: The power data of each mode of each fault node is imported into the local processor for processing and analysis. The specific processing process includes: Based on the proportional constraint factor after the first adjustment, the local processor adjusts the adoption ratio of each modal power data, extracts abnormal current data and abnormal voltage data from it, determines the abnormal type of each fault node, and matches the fault type of each fault node with the fault factor corresponding to each fault type pre-stored in the local processor to obtain the fault factor of each fault node. The fault factor is used to characterize the severity of the fault type. Based on the fault factor of each fault node, extract the first monitoring adjustment parameter corresponding to the fault factor, obtain the first monitoring adjustment parameter of each fault node, and adjust the monitoring period and data extraction frequency of each fault node; The first monitoring adjustment parameter of each fault node includes a monitoring period length adjustment parameter and a data extraction frequency adjustment parameter.

7. The multimodal power data analysis method based on the AI large model according to claim 1 is characterized by: If no diffusion fault occurs, the faulty device is identified, the data mode to which the faulty device belongs is obtained, the number of faulty devices in each mode of power data is counted, and the proportional constraint factor of each mode of power data is adjusted for a second time, specifically including: Collect statistics on the real-time performance status parameters and historical usage records of each device at a single fault node; The real-time performance status parameters include real-time load current, real-time vibration signal and harmonic content; The historical usage records include the cumulative length of time in use, the number of faults and the number of fault types; extracting real-time performance status verification parameters and historical usage record verification parameters from the built-in database of the data processing center; The real-time performance status verification parameters include the transformer real-time load verification current, real-time verification vibration signal and verification harmonic content; The historical usage record verification parameters include the cumulative usage time verification factor, the number of fault verifications and the number of fault type verifications; Based on the real-time performance status parameters and historical usage records of each device, the load index of each device is obtained after comparison and correction with the real-time performance status verification parameters and historical usage record verification parameters. The load index of each device is mapped and matched with the sinusoidal excitation current size corresponding to the load index in the built-in database of the data processing center to obtain the sinusoidal excitation current size of each device; Generates sinusoidal excitation current to each device and identifies the faulty device based on the response data of each device; Perform modal separation and feature extraction on the output data of each faulty device to obtain the main output mode of each faulty device and count the number of faulty devices in each modal power data; Based on the ratio of the number of faulty devices in each modal power data to the total number of faulty devices, a second adjustment is made to the proportional constraint factor of each modal power data in the abnormal period.

8. The multimodal power data analysis method based on the AI large model according to claim 1 is characterized by: The data is imported into the data processing center for analysis and processing to obtain the second monitoring adjustment parameter of the single fault node and adjust the monitoring period of the single fault node. The specific processing conditions are: Importing each modal power data into a data processing center, adjusting the adoption ratio of each modal power data based on a secondary adjustment ratio constraint factor, performing feature extraction on each modal power data, comprehensively evaluating the fault severity assessment value of a single fault node, mapping and matching the fault severity assessment value of the single fault node with the monitoring period adjustment parameters corresponding to the intervals of each fault severity assessment value pre-stored in a built-in database of the data processing center, obtaining the second monitoring period adjustment parameter of the single fault node, and monitoring the single fault node; The second monitoring period adjustment parameter of the single fault node includes a monitoring period length adjustment parameter and a data extraction frequency adjustment parameter.

9. A multimodal power data analysis system based on an AI large model, applying the multimodal power data analysis method based on an AI large model according to any one of claims 1 to 8, specifically comprising: The high-frequency modal monitoring module is used to input the real-time data stream from the power grid into the AI large model, monitor the high-frequency modes in the power grid in real time and identify abnormal events. Based on the data quality of each modal power data in the abnormal data stream, the proportional constraint factor of each modal power data is adjusted for the first time; The diffusion fault determination module is used to send the abnormal data stream to the data processing center to obtain the fault node, extract the modal power data of each adjacent node of the fault node for analysis, obtain the fault diagnosis results of each adjacent node, and determine whether a diffusion fault has occurred; A multi-node fault processing module is configured to, if a diffusion fault occurs, perform a hardware performance test on the built-in data processing module of each fault node based on the data volume of each modal power data of each fault node. If the hardware performance test result is qualified, the modal power data after adjusting the proportional constraint factor is input into the local processor under the built-in data processing module of each fault node for analysis to obtain the first monitoring adjustment parameter of each fault node, and adjust the monitoring period and data extraction frequency of each fault node; The single-node fault processing module is used to identify the faulty device if no diffusion fault occurs, obtain the data mode to which the faulty device belongs, count the number of faulty devices in each mode of power data, make a second adjustment to the proportional constraint factor of each mode of power data, and import it into the data processing center for analysis and processing to obtain the second monitoring adjustment parameter of the single fault node and adjust the monitoring period of the single fault node.

10. A multimodal power data analysis device based on an AI large model, applying the multimodal power data analysis method based on an AI large model according to any one of claims 1 to 8, characterized in that: include: The device has one or more programs, and the one or more programs are executed by one or more processors to implement the above method.

Images