Integrated current transformer fault diagnosis system and method based on machine learning

Abstract

The invention discloses an integrated current transformer fault diagnosis system and method based on machine learning, and relates to the technical field of fault diagnosis. Comprising a current transformer abnormity primary judgment module, a primary judgment abnormal current transformer state label judgment module, a disturbed current transformer disturbed data updating module and a fault source current transformer fault diagnosis module. According to the method, the feature vectors are constructed, and the multi-task gating neural network model is combined to carry out label probability determination, so that the body fault and the disturbed abnormity of the current transformer can be effectively distinguished, and corresponding diagnosis or data updating branches are automatically triggered for different fault types; under the complex working conditions of external electromagnetic interference, load fluctuation and the like, disturbed channel data are self-corrected in real time, so that the stability and the consistency of an overall data set are recovered, the misjudgment rate and the omission ratio are reduced, the robustness of transient disturbance and the fault positioning precision under a conjoined structure are improved, and the fault positioning efficiency is improved. And finally, reliable operation and long-term stability of the power grid measurement and protection system are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, in particular to a conjoined current transformer fault diagnosis system and method based on machine learning. BACKGROUND

[0002] As a key measurement and protection element in power systems, current transformers are widely used in substations, power transmission lines and industrial power distribution systems to transform high currents into measurable small currents in proportion, so as to realize the detection, measurement and protection control of current signals. Traditional current transformers are generally composed of primary winding, secondary winding, core and insulating shell, etc., and realize current conversion through electromagnetic induction principle. With the improvement of the automation level of power systems, the structural form of current transformers gradually diversifies. Among them, the conjoined current transformer gradually gets wide application in medium and low voltage distribution networks, switch cabinets and complete equipment due to its compact structure, convenient installation and strong adaptability.

[0003] In the prior art, the detection of the operating state of the current transformer mainly relies on manual inspection or traditional electrical parameter monitoring methods, by collecting parameters such as primary current, secondary current, temperature, insulation resistance, etc., to determine whether there is an abnormality. Some systems introduce online monitoring units and data acquisition modules, use the change trend of voltage, current and harmonic components to realize state recognition, and use signal filtering, threshold determination and feature extraction methods to assist state evaluation. With the development of smart grids, the state monitoring of current transformers gradually combines with digital and information technology. Existing researches propose to use signal feature analysis, time-frequency domain transformation, wavelet decomposition, Fourier analysis and other signal processing technologies to identify waveform distortion, phase shift and amplitude anomaly in the operation process of the current transformer. At the same time, some researches introduce pattern recognition and expert systems to build diagnostic models based on empirical rules to assist in determining fault types.

[0004] In the field of intelligent monitoring, machine learning methods have been applied to power equipment state evaluation and fault prediction. By extracting features and training patterns from a large amount of collected operating data, classification or regression models can be built to realize equipment state discrimination. For example, by using support vector machines, decision trees, neural networks and other algorithms, the working current, temperature, spectral features and load changes of the current transformer are learned and analyzed, so as to realize automatic state recognition and health evaluation to a certain extent. These technologies provide a basis for research and application reference for intelligent diagnosis of current transformers.

[0005] For example, the invention patent with the publication number CN117347801A discloses a combined mutual inductor insulation fault recognition method based on dynamic characteristics, which includes: 1) collecting voltage signals and current signals at the secondary side, and calculating the dynamic characteristics of the voltage signals and current signals; 2) comparing the dynamic characteristics with the operating data in the first database: for abnormal data, processing according to step 3); 3) comparing the abnormal data with the fault data in the second database: according to the comparison result, processing according to steps 4) and 5) respectively; 4) processing the abnormal data by using a pre-trained SVM model, identifying the fault type corresponding to the abnormal data and outputting; 5) manually identifying the abnormal data.

[0006] For example, the invention patent with the publication number CN117572140A discloses a current transformer secondary circuit neutral line grounding fault recognition method and device, which includes: obtaining a zero sequence current signal of a neutral point of a current transformer; decomposing the zero sequence current signal by using a VMD modulus decomposition algorithm to obtain a decomposed modal component; filtering out high-frequency noise modal components in the modal component to obtain a denoised modal component; analyzing and solving the denoised modal component based on a prony algorithm to obtain a harmonic component; and when the harmonic component meets a preset condition, determining that there is a grounding fault in the neutral line of the current transformer secondary circuit.

[0007] The prior art at least has the following technical problems:

[0008] A conjoined current transformer generally refers to a plurality of current transformer modules integrated in one body in the form of space, structure, electricity, etc., serving a multi-circuit measurement, protection or signal acquisition scene. Under this structure, each module not only has an independent electrical measurement function, but also has multiple coupling relationships such as space, magnetic field and circuit with other modules. The coupling fault propagation problem essentially means that when a local fault, degradation or abnormal behavior occurs in a module, this local problem is not completely confined within a single module, but may be transmitted or spread to other modules through various physical mechanisms, causing other modules without faults to also have performance deviation, measurement deviation or even indirectly inducing new abnormalities. SUMMARY

[0009] To solve the above technical problems in the prior art, the embodiments of the present application provide a conjoined current transformer fault diagnosis system and method based on machine learning. The technical solution is as follows:

[0010] On the one hand, a conjoined current transformer fault diagnosis system based on machine learning is provided, which includes:

[0011] The current transformer abnormality preliminary judgment module is configured to collect instantaneous current waveform data of each current transformer, calculate an abnormal offset data set of each current transformer, perform preliminary judgment of the abnormality of each current transformer, and record a current transformer with a suspected abnormality as a preliminary abnormality current transformer.

[0012] The preliminary abnormality current transformer state label judgment module is configured to construct a feature vector of each preliminary abnormality current transformer, input the feature vector into a machine learning model to obtain a label probability data set of the preliminary abnormality current transformer, analyze a state label of each preliminary abnormality current transformer, and determine a state execution branch of each preliminary abnormality current transformer.

[0013] The disturbed current transformer disturbed data update module is configured to record a corresponding preliminary abnormality current transformer as a disturbed current transformer when the state execution branch is a disturbed data update, determine a reference current transformer set, and perform disturbed data update of each disturbed current transformer.

[0014] The fault source current transformer fault diagnosis module is configured to record a corresponding preliminary abnormality current transformer as a source fault current transformer when the state execution branch is a fault diagnosis, and perform fault diagnosis of each source fault current transformer based on a machine learning model.

[0015] In another aspect, a machine learning-based coupled current transformer fault diagnosis method is provided, including the following steps:

[0016] The method includes collecting instantaneous current waveform data of each current transformer, calculating an abnormal offset data set of each current transformer, performing preliminary judgment of the abnormality of each current transformer, and recording a current transformer with a suspected abnormality as a preliminary abnormality current transformer.

[0017] The method includes constructing a feature vector of each preliminary abnormality current transformer, inputting the feature vector into a machine learning model to obtain a label probability data set of the preliminary abnormality current transformer, analyzing a state label of each preliminary abnormality current transformer, and determining a state execution branch of each preliminary abnormality current transformer.

[0018] The method includes recording a corresponding preliminary abnormality current transformer as a disturbed current transformer when the state execution branch is a disturbed data update, determining a reference current transformer set, and performing disturbed data update of each disturbed current transformer.

[0019] The method includes recording a corresponding preliminary abnormality current transformer as a source fault current transformer when the state execution branch is a fault diagnosis, and performing fault diagnosis of each source fault current transformer based on a machine learning model.

[0020] The technical scheme provided by the embodiment of the application has at least the following beneficial effects:

[0021] 1. The machine learning-based fault diagnosis system and method for a combined current transformer can achieve hierarchical identification and accurate tracing of abnormal signals in a multi-transformer combined operation environment. By constructing a feature vector and combining a multi-task gated neural network model for label probability determination, the body fault and disturbed abnormality of the current transformer can be effectively distinguished, and the corresponding diagnosis or data update branch can be automatically triggered for different fault categories. In the data update stage, the system introduces reference transformer similarity and multi-dimensional weighted smoothing mechanism, which can correct the disturbed channel data in real time under complex working conditions such as external electromagnetic interference and load fluctuation, thereby restoring the stability and consistency of the overall data set. The misjudgment rate and the missed detection rate are reduced, the typical fault modes such as inter-turn short circuit, core saturation and open circuit drift are identified, the robustness of the system to transient disturbance and the fault positioning accuracy under the combined structure are improved, and finally the reliable operation and long-term stability of the power grid measurement and protection system are ensured.

[0022] 2. The application can accurately capture early micro performance drift and nonlinear waveform distortion by performing abnormal preliminary judgment of each current transformer, automatically adjusting the abnormal judgment sensitivity for different operating conditions, and providing a data basis for subsequent state label discrimination.

[0023] 3. The application can realize intelligent shunting and accurate attribution of multi-source abnormalities of the combined current transformer by analyzing the state label of each preliminary abnormal current transformer and determining the state execution branch of each preliminary abnormal current transformer. The misdiagnosis or repeated calculation caused by category confusion is avoided, the source fault transformer is preferentially triggered for deep diagnosis and analysis, the fault is identified, and the reference data compensation and signal smoothing of the disturbed transformer are performed to eliminate false abnormalities caused by external factors such as electromagnetic interference and load mutation. The classification decision accuracy and response efficiency of the system in complex electromagnetic environment are improved, and dynamic autonomy of the diagnosis path and hierarchical closed-loop control of fault processing are realized. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0025] Figure 1 is a structural diagram of the machine learning-based fault diagnosis system for a combined current transformer provided by the embodiments of the application.

[0026] Figure 2Is the embodiment of the present application provides a flow chart of the fault diagnosis method of the combined current transformer based on machine learning.

[0027] Figure 3 Is the state label judgment and execution branch determination flow chart related to the embodiment of the present application.

[0028] Figure 4 Is the disturbed data update and fault diagnosis comprehensive flow chart related to the embodiment. DETAILED DESCRIPTION

[0029] The technical solutions in the present application will be described below with reference to the drawings.

[0030] In the embodiments of the present application, the words such as "example", "for example" are used to represent as an example, illustration or explanation. Any embodiment or design scheme described as "example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. In fact, the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or one of the two.

[0031] In the embodiments of the present application, "image" and "picture" can be used interchangeably sometimes. It should be pointed out that when the distinction is not emphasized, the meanings expressed are consistent. "Of", "corresponding" and "corresponding" can be used interchangeably sometimes. It should be pointed out that when the distinction is not emphasized, the meanings expressed are consistent.

[0032] In the embodiments of the present application, sometimes the subscript such as W1 can be written in the form of non-subscript such as W1. When the distinction is not emphasized, the meanings expressed are consistent.

[0033] In order to make the technical problems, technical solutions and advantages to be solved by the present application clearer, specific embodiments will be described in detail below with reference to the drawings.

[0034] The embodiments of the present application provide a machine learning-based combined current transformer fault diagnosis system and method, as shown in Figure 1 The structure schematic diagram of the machine learning-based combined current transformer fault diagnosis system, including: current transformer abnormality preliminary judgment module, preliminary judgment abnormal current transformer state label judgment module, disturbed current transformer disturbed data update module and fault source current transformer fault diagnosis module.

[0035] The current transformer abnormality preliminary judgment module is used for collecting the instantaneous current waveform data of each current transformer, calculating the abnormal offset data set of each current transformer, and performing the abnormality preliminary judgment of each current transformer, and recording the current transformer with the abnormality preliminary judgment result as a suspected abnormality as a preliminary judgment abnormality current transformer.

[0036] Further, the instantaneous current waveform data of each current transformer is collected, and the abnormal offset data set of each current transformer is calculated, and the specific calculation process is as follows:

[0037] The sampling period is set, the high-frequency sampling is performed on the secondary side of each current transformer, and the instantaneous current waveform data of each current transformer is obtained.

[0038] In the embodiment of the present application, in order to realize accurate analysis of the harmonic component and waveform characteristics in the secondary side output signal of the conjoined current transformer, the system pre-sets the sampling period parameter suitable for high-frequency waveform data collection. The selection principle of the sampling period is based on the Nyquist sampling theorem, that is, the sampling frequency needs to be not less than twice the highest frequency component of the measured signal. Considering that the upper limit of the common harmonic analysis in the power system is usually 25 or 40 (based on 50Hz power frequency signal), the highest frequency component can reach 1250Hz or 2000Hz. Therefore, under the premise of ensuring complete capture of harmonics, the present application preferably sets the sampling frequency to be not less than 10kHz, corresponding to a sampling period of not more than 100 microseconds (i.e. 0.1ms). This sampling frequency is sufficient to cover the main harmonic components in the conventional power system, and also takes into account the fine depiction of high-frequency interference or fault characteristics, ensuring high-fidelity collection of current signals. In addition, the specific value of the sampling period can also be dynamically adjusted according to the system bandwidth, data processing capacity and actual operation scene. When the system detects a waveform mutation, abnormal offset or the upper computer issues an analysis request, the sampling frequency can be temporarily increased to a higher gear (e.g. above 20kHz) to obtain higher resolution transient information for further fault mode identification and multi-modal discriminant analysis, thereby meeting the accuracy and timeliness requirements of the conjoined current transformer state monitoring.

[0039] A current transformer (CT) is an electromagnetic induction device used for measuring alternating current or providing signals for a relay protection system, widely used in power systems, electrical equipment and industrial control systems. Its core function is to transform high current into smaller standard current (such as 5A or 1A) according to a certain proportion, facilitating the access of measurement and protection equipment, while ensuring the safety of the measurement system.

[0040] A current transformer generally consists of the following parts: primary winding: directly connected in series with the measured circuit, bearing the original current, usually with fewer turns (usually 1 turn). Secondary winding: connected to measuring instruments or protective relays, with many turns, used to output scaled-down current. Core: provides a magnetic flux path to induce the primary side current to the secondary side current. Insulating shell and terminals: ensure safety and stability. Its working principle is based on the law of electromagnetic induction, the primary current generates alternating magnetic flux in the core, and the secondary winding induces a current proportional to the primary current.

[0041] A combined current transformer (Combined Current Transformer or Multi-in-one CT) is a current transformer structure that integrates multiple measurement channels or multiple phases (such as three-phase) in the same structure shell, with the characteristics of compact structure, simple wiring, strong synchronization, etc. It is commonly used in low-voltage switchgear, feeder terminal, ring network cabinet, intelligent meter box and distribution automation system. Each channel still operates according to the principle of traditional CT (electromagnetic induction), and its essence is multiple CTs in parallel, but in physical structure, they share a shell, a base, and possibly a core (part of the structure). Electromagnetic characteristics are independent of each other in principle, but may experience cross-talk in cases of magnetic interference and spatial coupling.

[0042] It needs to be explained that the instantaneous current waveform data refers to the sequence of current values that vary with time continuously collected from the secondary side output of the current transformer at a set sampling period. This sequence truly depicts the instantaneous size and fluctuation characteristics of the current at each time point.

[0043] The current effective value of each current transformer is calculated based on the instantaneous current waveform data of each current transformer.

[0044] In a specific embodiment, for a current transformer, the current effective value is defined as: where I RMS is the current effective value, i k is the instantaneous current at the kth sampling time point, k = 1, 2,..., N, k is the number of sampling time points, and N is the total number of sampling time points.

[0045] It also needs to be added that in another specific embodiment, if the current signal can be represented as a continuous-time function i(t) in a complete period, and its period is T, then the current effective value I RMS can be defined based on the integral form, specifically: .

[0046] It needs to be explained that compared with the root mean square calculation method based on discrete sampling points, the continuous integral formula is closer to the theoretical definition and can provide more ideal accuracy in high sampling frequency or analog modeling. However, in engineering practice, due to real-time and calculation efficiency, the discrete form is often preferred for current effective value estimation.

[0047] Synchronously based on the instantaneous current waveform data of each current transformer, frequency domain transformation and fundamental wave extraction are performed to calculate the total harmonic distortion rate of each current transformer.

[0048] In specific embodiments, for a current transformer, the instantaneous current waveform data of the current transformer is subjected to fast Fourier transform to obtain the corresponding current frequency spectrum distribution, and thus the fundamental frequency (such as 50 Hz) and the high-order harmonic components (such as 150 Hz, 250 Hz, etc.) are identified by spectral analysis, which correspond to the current amplitudes of different frequency components, respectively. The amplitude information of each frequency component in the frequency spectrum is extracted, the fundamental frequency is set as a reference, and the amplitudes of the remaining high-order harmonic components are denoted as I2, I3,..., In, respectively. n where n is the highest harmonic order, and thus the total harmonic distortion rate (THD) is defined as: .

[0049] The historical data set of each current transformer is obtained, and the historical characteristic data of each current transformer is calculated, including the historical current effective value mean, the historical current effective value standard deviation, the historical average total harmonic distortion rate, and the historical total harmonic distortion rate standard deviation.

[0050] In specific embodiments, based on the set historical statistical period length (set according to experience, common configurations include 24 hours, 48 hours, 72 hours, or nearly 7 days), the historical data set of each current transformer is extracted from the edge database, including the historical current effective value sequence and the historical total harmonic distortion rate sequence of each current transformer.

[0051] The abnormal offset data set of each current transformer is calculated based on the current effective value of each current transformer, the total harmonic distortion rate of each current transformer, and the historical characteristic data of each current transformer.

[0052] The abnormal offset data set of each current transformer includes the current effective value abnormal offset degree of each current transformer and the total harmonic distortion rate abnormal offset degree of each current transformer.

[0053] In specific embodiments, for a current transformer, the current effective value abnormal offset degree is defined as: where δ I is the current effective value abnormal offset degree, I RMS is the current effective value, is the historical current effective value mean, and σ Iis the standard deviation of the historical current effective value.

[0054] The abnormal offset degree of the total harmonic distortion rate is defined as: wherein, δ THD is the abnormal offset degree of the total harmonic distortion rate, THD is the total harmonic distortion rate, is the historical average total harmonic distortion rate, σ THD is the standard deviation of the historical total harmonic distortion rate.

[0055] Further, the abnormal preliminary judgment of each current transformer is performed, and the specific implementation process is as follows:

[0056] The abnormal density of each current transformer in the preset historical statistical duration is calculated and is recorded as the historical abnormal density of each current transformer.

[0057] In the specific embodiment of the present application, in order to improve the accuracy and adaptive ability of the abnormal judgment of the current transformer, the historical abnormal density is introduced as the dynamic adjustment data basis of the offset threshold of each current transformer. The historical abnormal density is used to quantify the proportion of abnormal behavior of the current transformer in a historical operation period. The historical analysis period, which is in hours or days, commonly takes values including 24 hours, 48 hours, 72 hours or nearly 7 days, and the specific value is set according to the comprehensive setting of the operation mode of the power system, the periodicity of the load and the frequency of data fluctuation. In order to enhance the adaptability and real-time performance of the system, the historical analysis period is refreshed periodically in combination with the sliding window mechanism, and the sliding step can be set to be different from 30 minutes to 1 hour. Assuming that the historical analysis period is set to 48 hours, the sampling period of the system is 5 minutes, and there are 576 periods in total, if 29 periods are judged to be abnormal in this period, then the historical abnormal density is 29 divided by 576, which is about 0.05.

[0058] The deviation threshold correction coefficient of each current transformer is extracted based on the historical abnormal density of each current transformer.

[0059] In the embodiment of the present application, the system pre-constructs a mapping set of abnormal density and deviation threshold correction coefficient. The mapping set can adopt a discrete segmented form, for example, the abnormal density is divided into several intervals (such as <1%, 1-3%, 3-6%, 6-10%, >10%), which respectively correspond to correction coefficients (such as 1.10, 1.00, 0.90, 0.80, 0.70), or a continuous mapping relationship using function fitting. In the actual application process, the historical abnormal density is used as an index to extract the deviation threshold correction coefficient from the mapping set, so as to realize the differential judgment of different abnormal sensitivity modules and improve the accuracy and adaptive ability of abnormal detection. The mapping set supports dynamic optimization and online adjustment, and can be continuously updated according to large-scale operation data to adapt to the evolution of the actual scene.

[0060] The current effective value abnormal offset correction threshold of each current transformer is obtained by multiplying the current effective value abnormal offset threshold by the deviation threshold correction coefficient of each current transformer, and the total harmonic distortion rate abnormal offset correction threshold of each current transformer is obtained by multiplying the total harmonic distortion rate abnormal offset threshold by the deviation threshold correction coefficient of each current transformer.

[0061] If the current effective value abnormal offset degree of a certain current transformer exceeds the current effective value abnormal offset correction threshold of the current transformer or the total harmonic distortion rate abnormal offset degree exceeds the total harmonic distortion rate abnormal offset correction threshold of the current transformer, it indicates that the electrical characteristics of the current transformer in the current period deviate from its long-term behavior mode, and there is a potential operation abnormality. The abnormal preliminary judgment result of the current transformer is suspected abnormal, and the current transformer is recorded as a preliminary judgment abnormal current transformer.

[0062] If the current effective value abnormal offset degree of a certain current transformer does not exceed the current effective value abnormal offset correction threshold of the current transformer and the total harmonic distortion rate abnormal offset degree does not exceed the total harmonic distortion rate abnormal offset correction threshold of the current transformer, it indicates that the current running state of the current transformer is highly consistent with its historical running state, and no statistically significant abnormality is shown. The abnormal preliminary judgment result of the current transformer is normal, and the current transformer is recorded as a preliminary judgment normal current transformer.

[0063] The preliminary judgment abnormal current transformer state label judgment module is used to construct the feature vector of each preliminary judgment abnormal current transformer, and outputs the label probability data set of the preliminary judgment abnormal current transformer based on the machine learning model, analyzes the state label of each preliminary judgment abnormal current transformer, and determines the state execution branch of each preliminary judgment abnormal current transformer.

[0064] As Figure 3As shown, it is the state label judgment and execution branch determination flow chart related to the embodiment of the application. After completing the initial judgment of each current transformer, the feature vector of each initial abnormal current transformer is obtained, the feature vector of each initial abnormal current transformer is input into the pre-trained machine learning model, and the label probability data set of each initial abnormal current transformer is output by the machine learning model. The state label of each initial abnormal current transformer is determined based on the maximum value of the probability in the label probability data set. If the state label is the source fault label, enter the state execution branch source fault diagnosis, and execute the source fault related diagnosis process; if the state label is the disturbed label, enter the state execution branch disturbed data update, update the disturbed state related data and perform dynamic correction; if the state label is the normal label, enter the state execution branch continue sampling detection, and continue to execute the periodic sampling and monitoring of the current transformer.

[0065] Further, the feature vector of each initial abnormal current transformer is constructed, and the specific construction process is as follows:

[0066] The number of each initial abnormal current transformer is obtained.

[0067] The first abnormal time stamp of each initial abnormal current transformer is extracted, and thus the earliest abnormal judgment coefficient of the initial abnormal current transformer with the earliest abnormality is set to 1, and the earliest abnormal judgment coefficient of the remaining initial abnormal current transformer is set to 0.

[0068] In a specific embodiment, when the current transformer is determined to be initially abnormal, the system records the current time point as the first abnormal time stamp of the current transformer. The time stamp is stored in the event record table or the edge time series database as an important basis for subsequent traceability judgment. In order to further clarify the source fault identification logic in the multi-module abnormal scene, the system assigns a higher source fault indication weight to the current transformer that first occurs abnormality. For this purpose, the earliest abnormal judgment coefficient is set as an input in the feature vector: if the abnormal time stamp of a certain initial abnormal current transformer is the earliest among all initial abnormal current transformers, the earliest abnormal judgment coefficient of the current transformer is set to 1; the earliest abnormal judgment coefficient of the remaining initial abnormal current transformer is set to 0. The setting of the earliest abnormal judgment coefficient helps to improve the accuracy of the machine learning model (a multi-task MLP model is used in this embodiment) in the source fault module determination. Since the source fault module is often the first to be physically disturbed (such as inter-turn short circuit, insulation breakdown, core saturation, etc.), its abnormal characteristic data will appear significant deviation in time sequence before the disturbed module. Therefore, the introduction of the earliest abnormal judgment coefficient can effectively assist the machine learning model to distinguish between the active fault source and the passive disturbed object, thereby improving the prediction probability confidence of the machine learning model for the source fault module class and optimizing the execution efficiency of the fault alarm and disposal strategy.

[0069] The electromagnetic interference intensity of each initial abnormality current transformer is extracted based on the first abnormality time stamp of each initial abnormality current transformer.

[0070] In specific embodiments, to extract the electromagnetic interference intensity of each initial abnormality current transformer before the occurrence of an abnormality, first, the first abnormality time stamp of each initial abnormality current transformer obtained in the foregoing step is used as a positioning key point. The electromagnetic field monitoring data stream recorded in the edge database is called, which can be collected in real time by electromagnetic interference sensors (such as Hall effect sensors, fluxgate sensors, etc.) deployed between devices or around the bus, and recorded at high frequency according to a unified time stamp. To ensure that the obtained electromagnetic interference intensity is targeted and timely, the system takes the previous sampling period corresponding to the first abnormality time stamp as the reference window, indexes the original electromagnetic intensity value at that time from the historical electromagnetic interference data, and uses it as the electromagnetic interference intensity representation value of the initial abnormality current transformer. The specific extraction process is as follows: the system backtracks one sampling period according to the first abnormality time stamp of the current transformer; queries the electromagnetic intensity value at the corresponding time in the electromagnetic interference data stream; uses the electromagnetic intensity value as the electromagnetic interference intensity of the initial abnormality current transformer at the moment before the first abnormality, which is used to construct the input feature vector of the subsequent machine learning model. This helps to identify false abnormality situations caused by external magnetic field interference, thereby improving the ability to distinguish disturbed modules and source fault modules in the machine learning model, suppressing the false positive rate, and enhancing the environmental robustness of the system.

[0071] The bus voltage is collected, and the current-bus voltage correlation coefficient of each initial abnormality current transformer is calculated.

[0072] In specific embodiments, to improve the accuracy of abnormality determination, bus voltage data is further collected, and the current-bus voltage correlation coefficient of each initial abnormality current transformer is calculated to identify whether there is a non-organic abnormality caused by system voltage fluctuation.

[0073] For a primary abnormality judgment current transformer, first, a voltage sensor such as a voltage transformer (VT) is arranged on the low-voltage bus, and the instantaneous waveform data of the bus voltage is synchronously collected at a set sampling period (such as 10 ms), and is time-aligned and window-paired with the current effective value data of the secondary side of the primary abnormality judgment current transformer. This process ensures that the voltage-current data pair has synchronization and comparability in the same time period. Then, the numbers of the system primary abnormality judgment current transformers are extracted, and the current effective value sequences and the corresponding bus voltage effective value sequences in the preset time period are extracted. The Pearson correlation coefficient is used to analyze the correlation of the above two time sequences, and the current-bus voltage correlation coefficient of the primary abnormality judgment current transformer is obtained.

[0074] The current-bus voltage correlation coefficient of each primary abnormality judgment current transformer represents the response degree of the current change of each current transformer to the bus voltage fluctuation. If the current abnormality of a certain primary abnormality judgment current transformer is highly correlated with the voltage fluctuation, it means that the abnormality may be caused by system disturbance, and does not have independent fault characteristics; on the contrary, it may be a fault signal of the module itself.

[0075] The abnormal offset data set of each primary abnormality judgment current transformer is extracted from the abnormal offset data set of each current transformer.

[0076] The numbers of the primary abnormality judgment current transformers, the earliest abnormality judgment coefficients of the primary abnormality judgment current transformers, the electromagnetic interference intensities of the primary abnormality judgment current transformers, the current-bus voltage correlation coefficients of the primary abnormality judgment current transformers, and the abnormal offset data sets of the primary abnormality judgment current transformers are jointly used as the state label analysis feature vectors of the primary abnormality judgment current transformers.

[0077] The instantaneous waveform signal features, the environmental state context features, and the device historical abnormality statistical features of each primary abnormality judgment current transformer are collected and analyzed, and are jointly used as the supplementary feature vectors of each primary abnormality judgment current transformer.

[0078] In this embodiment, the transient waveform signal feature refers to the data obtained by sampling the high-frequency current waveform of the secondary side of each initially abnormal current transformer based on a set sampling period, and the waveform shape parameters extracted through time domain and frequency domain analysis, specifically including current waveform peak value, peak-to-peak value, skewness, kurtosis, harmonic energy distribution of each order, fundamental frequency offset, etc. The sampling process can use an ADC module with a sampling rate of ≥10 kHz, and combine fast Fourier transform (FFT) or wavelet analysis method for signal feature extraction. The environmental state context feature is used to reflect the power distribution topology and operating environment conditions of the current transformer, and has an auxiliary role in the attribution of abnormal behavior. Such features include: the location topology information of the initially abnormal current transformer (such as bus segment, feeder number), the corresponding loop type (single phase / three phase, important / ordinary load), the effective value of bus voltage at the collection period, the current voltage correlation coefficient, the system total load fluctuation rate, the adjacent current transformer state, the field electromagnetic interference intensity (which can be collected by an EMI sensor), and the temperature and humidity environment parameters, etc. The above information can be obtained from the SCADA system, edge node sensors, and EMI monitoring module. The device historical abnormal statistical feature is used to identify whether the current transformer has a long-term degradation or frequent abnormal trend. Based on a set historical statistical period (preset according to experience, such as 7 days, 30 days, 60 days), it extracts the historical current effective value sequence and the historical total harmonic distortion rate sequence of the initially abnormal current transformer from the edge database, and further calculates the corresponding abnormal frequency, abnormal density, continuous abnormal period number, time interval from the last abnormality, etc. statistical indicators.

[0079] The state label analysis feature vector of each initially abnormal current transformer and the augmented feature vector of each initially abnormal current transformer are combined as the feature vector of each initially abnormal current transformer.

[0080] Further, the label probability dataset of the initially abnormal current transformer is output, and the specific output process is:

[0081] The feature vector of each initially abnormal current transformer is input into the pre-trained machine learning model, and the machine learning model adopts a multi-task MLP gating model, which outputs the label probability dataset of each initially abnormal current transformer through a main discrimination channel.

[0082] The label probability dataset of each initially abnormal current transformer includes the fault source probability, the disturbed probability and the normal probability of each initially abnormal current transformer.

[0083] In the embodiment, the multi-task MLP gated model is a multi-task learning model with a main discrimination channel and a sub-task diagnosis channel, adopts a multi-layer perception network structure, and fuses a gating mechanism to realize dynamic switching control of tasks. The main structure of the model takes a multi-layer fully connected neural network as a core, completes label probability output of the primary abnormal current transformer through the main discrimination channel, specifically includes three types of labels of fault source probability, disturbed probability and normal probability, which are used as the basis for type judgment in the subsequent processing process. The main discrimination channel converts the probability distribution of the model output through a normalization activation function, so as to ensure that the output result has explainability and clear discrimination boundary. The above model is pre-trained based on a historical labeled data set in the training stage. The training data includes the earliest abnormal judgment coefficient of the current transformer, the electromagnetic interference intensity, the current-bus voltage correlation coefficient, the abnormal offset data set, the instantaneous waveform signal feature, the environmental state context feature and the device historical abnormal statistical feature, and the corresponding current transformer state label and fault type label. In the training process, the main task and the sub-task are optimized in parallel through a joint loss function. The main discrimination channel is used to maximize the type recognition accuracy, and the sub-task diagnosis channel is used to improve the fine-grained fault discrimination ability of the source fault transformer, so as to finally form a double-channel structure model with type judgment and diagnosis ability. In the inference stage, the model running process is sequentially executed. First, the feature vector of each primary abnormal current transformer is input into the main discrimination channel of the model to obtain the corresponding three types of label probability values. Then, whether to activate the fault diagnosis module is determined based on the gating mechanism. If the fault source probability of a current transformer is the largest, the fault diagnosis sub-channel is automatically entered to perform further diagnosis analysis. Otherwise, the inference result is kept as disturbed or normal state.

[0084] Further, the state labels of each primary abnormal current transformer are analyzed to determine the state execution branch of each primary abnormal current transformer. The specific process is as follows:

[0085] Based on the label probability data set of each primary abnormal current transformer, the probability type with the largest probability is recorded as the state label of the corresponding primary abnormal current transformer.

[0086] The state label includes a source fault label, a disturbed label and a normal label.

[0087] According to the state label, each primary abnormal current transformer is divided into each source fault current transformer, a disturbed current transformer and a normal current transformer.

[0088] If a certain initially abnormal current transformer is a disturbed current transformer, it means that its disturbed label probability is the largest, i.e. higher than the source fault probability and the normal probability, so that the abnormal performance of the initially abnormal current transformer is mainly from external electromagnetic interference, environmental noise, upstream system fluctuation and other non-own fault reasons, rather than its own performance abnormality. The state execution branch of the initially abnormal current transformer is recorded as disturbed data updating.

[0089] It needs to be explained that in the actual application scenario where the embodiment is located, such as an intelligent substation or a power system state monitoring platform, if the initially abnormal current transformer is identified as a disturbed current transformer, it means that the signal abnormality of the initially abnormal current transformer is mainly caused by external non-own interference, such as electromagnetic interference, environmental noise or upstream system fluctuation and other non-device fault factors. Although these abnormalities are not caused by the performance degradation of the device itself, the waveform, current value or harmonic characteristic may still deviate from the normal range. If the abnormality is not processed and directly involved in subsequent analysis, it will mislead the system overall state judgment and device behavior reasoning. For example, it may misjudge that the system load is too high, mis-trigger the protection strategy, interfere with the health prediction model based on data driving, and thus affect the safe operation of the power grid or the precision of device operation and maintenance. Therefore, in order to avoid the misleading effect of the disturbed data, the system needs to perform a disturbed data updating process on the disturbed current transformer, i.e. taking the normal current transformer as a reference, correcting its key feature signals (such as current effective value, total harmonic distortion rate, current phase angle, power factor, etc.) through feature similarity matching and weighted smoothing compensation, so as to restore the reference value and physical consistency of the data. Such data updating is not for correcting model misjudgment, but for compensatory correction of the quality of the system data source, to ensure the logical closed loop and stable output of the overall analysis chain. This mechanism is particularly important in a multi-sensor redundant structure, which can improve the fault tolerance and analysis reliability of the system to environmental interference.

[0090] If a certain initially abnormal current transformer is a source fault current transformer, it means that the abnormal mode of the initially abnormal current transformer is closest to the representation of the fault current transformer in the training samples, and has an independent fault risk. The state execution branch of the initially abnormal current transformer is recorded as source fault diagnosis.

[0091] If a certain initially abnormal current transformer is a normal current transformer, it means that the current features of the initially abnormal current transformer are within an acceptable range, and there is no need to be classified as a fault or a disturbance. The initially abnormal current transformer is considered to have no abnormal performance temporarily, and no alarm or processing is triggered. The state execution branch of the initially abnormal current transformer is recorded as continuing sampling detection, which specifically includes: maintaining the normal high-frequency sampling and data storage; re-evaluating the fluctuation trend of the initially abnormal current transformer in the next diagnosis cycle; and if subsequent abnormalities occur, the initially abnormal current transformer can re-enter the initial abnormality judgment process.

[0092] For example, Figure 4The figure is a comprehensive flowchart of the disturbed data updating and fault diagnosis involved in the embodiment. After completing the state label determination of each preliminary abnormal current transformer, enter the corresponding state execution branch according to the label result. When the judgment result is disturbed data updating, the system first determines the reference current transformer set and judges whether there is a preliminary normal or normal current transformer. If not, generate a prompt message and do not perform data updating; if so, calculate the similarity between each basic reference current transformer and the disturbed current transformer. The system further extracts the abnormal offset data set and calculates the similarity replacement score. If there is a similarity replacement score greater than or equal to the similarity replacement score threshold, the reference current transformer set is determined. Then, the system calculates the multi-dimensional compensation data and performs exponential smoothing update to realize dynamic correction and update of the multi-dimensional data of the disturbed current transformer. When the judgment result is source fault diagnosis, the system triggers the gate mechanism to start the fault diagnosis sub-network, reuses the input feature vector of the source fault current transformer, performs fault pattern recognition analysis based on the fault diagnosis sub-network, and finally outputs the fault type and fault level result of the source fault current transformer.

[0093] The disturbed data updating module of the disturbed current transformer is used to record the corresponding preliminary abnormal current transformer as the disturbed current transformer when the state execution branch is disturbed data updating, determine the reference current transformer set, and thereby perform the disturbed data updating of each disturbed current transformer.

[0094] Further, the reference current transformer set is determined, and the specific determination process is as follows:

[0095] For a disturbed current transformer, if there is no preliminary normal current transformer and normal current transformer, generate a prompt message and do not perform disturbed data updating.

[0096] It should be noted that when a preliminary abnormal current transformer is determined to be a disturbed current transformer, the subsequent disturbed data updating steps depend on the available normal reference object, that is, the preliminary normal current transformer and the normal current transformer. If both do not exist, the system cannot correct or replace the disturbed data, at which time a prompt message is generated and updating is stopped.

[0097] If there is a preliminary normal current transformer or a normal current transformer, the preliminary normal current transformer or the normal current transformer is jointly recorded as each basic reference current transformer.

[0098] Based on each basic reference current transformer, the basic reference-disturbed current transformer similarity is calculated.

[0099] In specific embodiments, each basic reference-disturbed current transformer similarity reflects the similarity between the disturbed current transformer and each basic reference current transformer in the operating state, and the core goal is to select the reference object that best represents the true state of the current disturbed transformer for subsequent compensation correction. Specifically, the state feature vector of the disturbed current transformer contains multiple key indicators representing the operating conditions and current quality, such as current effective value (Irms), current phase angle (Φ), power factor (PF), and total harmonic distortion rate (THDi), etc. For each pair of disturbed-basic reference current transformers, first, the above features are extracted to form a multi-dimensional state vector, and each component of the vector is normalized to eliminate the bias caused by the dimension and scale differences of the index. Then, a multi-dimensional similarity calculation method is used to compare the transformers. Common calculation methods include Euclidean distance and cosine similarity: Euclidean distance reflects the absolute difference in the feature space, which is suitable for scenarios where the feature scales are consistent; while cosine similarity focuses on the consistency of the direction of the state vector, which is more suitable for situations where the amplitude fluctuation is obvious but the feature structure is similar. In most power systems, to improve the robustness of the model to load amplitude fluctuations, it is recommended to use weighted cosine similarity, and different features are assigned weights according to their impact on system stability, such as assigning a higher weight to THDi to enhance the recognition ability of harmonic disturbances.

[0100] Based on the abnormal offset data set of each current transformer, the abnormal offset data set of each basic reference current transformer is extracted, including the current effective value abnormal offset degree of each basic reference current transformer and the total harmonic distortion rate abnormal offset degree of each basic reference current transformer.

[0101] The reference basic reference-disturbed current transformer similarity, the reference current effective value abnormal offset degree, and the reference total harmonic distortion rate abnormal offset degree preset in the database are extracted.

[0102] The basic reference-disturbed current transformer similarity weighting factor, the current effective value abnormal offset degree weighting factor, and the total harmonic distortion rate abnormal offset degree weighting factor preset in the database are extracted.

[0103] In the embodiment, the system database is preconfigured with a mapping relationship of different characteristic dimension weighting factors for reference object screening of the disturbed current transformer, which is used to quantify the weight distribution of the basic reference-disturbed current transformer similarity, the current effective value abnormal offset degree, and the total harmonic distortion rate abnormal offset degree in the similar substitution score calculation. The mapping relationship is managed in the form of a structured weighting factor table, which is constructed by combining a large amount of coupled current transformer operation data, historical disturbed cases, and fault diagnosis verification results. The influence of different characteristic dimensions on the accuracy of disturbed state discrimination is statistically analyzed to determine the basic weight of each dimension, and the basic weight is fixed in the database after multiple iterations and optimization. Based on the mapping system, the system can directly extract the basic reference-disturbed current transformer similarity weighting factor, the current effective value abnormal offset degree weighting factor, and the total harmonic distortion rate abnormal offset degree weighting factor from the database. Each weighting factor is a real number, and the range is limited to 0 to 1, and the sum of the three is equal to 1, which ensures that the influence weight of different characteristic indicators in the coupled calculation of the similar substitution score conforms to the discrimination logic in the actual operation scene, avoids the over-dominance of a single characteristic in the screening result, and improves the scientificity and reliability of the reference current transformer selection.

[0104] The basic reference-disturbed current transformer similarity is compared with the reference basic reference-disturbed current transformer similarity, the reference current effective value abnormal offset degree is compared with the current effective value abnormal offset degree of each basic reference current transformer, and the reference total harmonic distortion rate abnormal offset degree is compared with the total harmonic distortion rate abnormal offset degree of each basic reference current transformer. The coupling processing is performed on the comparison processing results combined with the weighting factors to obtain the similar substitution scores of each basic reference current transformer.

[0105] In specific embodiments, the similar substitution score of a basic reference current transformer is defined as:

[0106] ,

[0107] Wherein, A is the similar substitution score of the basic reference current transformer, a is the basic reference-disturbed current transformer similarity, b is the current effective value abnormal offset degree of the basic reference current transformer, c is the total harmonic distortion rate abnormal offset degree of the basic reference current transformer, a0 is the reference basic reference-disturbed current transformer similarity, b0 is the reference current effective value abnormal offset degree, c0 is the reference total harmonic distortion rate abnormal offset degree, x1 is the basic reference-disturbed current transformer similarity weighting factor, x2 is the current effective value abnormal offset degree weighting factor, and x3 is the total harmonic distortion rate abnormal offset degree weighting factor.

[0108] The similar substitution score threshold value preconfigured in the database is extracted.

[0109] Before performing the data update of the disturbed current transformer, to ensure that the reference object has sufficient representativeness and substitution ability, the system needs to extract a preset similar substitution score threshold from the database. This threshold is used to evaluate the similar substitutability of each basic reference current transformer for the disturbed current transformer, that is, whether it has the reference qualification as the basis for data update. The setting of the similar substitution score threshold is based on historical statistical analysis and engineering experience in multiple dimensions, including: historical substitution effect evaluation (based on a large amount of historical operation data, the update result deviation under different similar substitution score values and the misjudgment rate after repair are counted, and the score corresponding to the error control within the engineering tolerance range (such as within 5%) is selected as the threshold) and expert experience modeling (combining the cognition of system experts on device working conditions, the influence of each feature on device performance stability is weighted, and the score threshold is determined in the experimental tuning process).

[0110] If there is no basic reference current transformer with a similar substitution score greater than or equal to the similar substitution score threshold, it means that there is no sensor in the current system that has sufficient matching in the running state with the disturbed current transformer. At this time, forcibly using these non-representative objects for data correction may introduce new errors or mislead subsequent system judgment, so a prompt information is generated and the disturbed data update is not performed.

[0111] If there is a basic reference current transformer with a similar substitution score greater than or equal to the similar substitution score threshold, it means that there is a basic reference current transformer that has sufficient consistency in running characteristics with the disturbed current transformer, which has the substitutability as a reference. The corresponding basic reference current transformer is recorded as the reference current transformer.

[0112] Traverse each reference current transformer to obtain a set of reference current transformers.

[0113] Traverse each disturbed current transformer to obtain a set of reference current transformers corresponding to each disturbed current transformer.

[0114] Further, the disturbed data update of each disturbed current transformer is performed, and the specific analysis process is as follows:

[0115] For a disturbed current transformer, obtain the set of reference current transformers corresponding to the disturbed current transformer, and thereby obtain a set of multi-dimensional compensation data of each reference current transformer.

[0116] In specific embodiments, the set of multi-dimensional compensation data mainly includes the following key feature dimensions: current effective value, total harmonic distortion rate, current phase angle, and power factor.

[0117] Among them, the current effective value (RMS) represents the current amplitude intensity per unit time, which is used to measure the load level of the current working condition and constitutes the basis index for the correction of the disturbed data current amplitude. The total harmonic distortion rate (THD) reflects the spectral purity of the current signal on the branch where the mutual inductor is located, which can be used to identify whether there is non-power frequency disturbance or harmonic interference, and is an important reference for data filtering and abnormality judgment. The current phase angle refers to the phase difference between the current and the system reference voltage, which is used to judge the change of system power factor and load property characteristics, and to assist in identifying whether the deviation source comes from the load nonlinearity. The power factor represents the ratio of active power to apparent power, which directly reflects the energy conversion efficiency of the system and can assist in identifying power end disturbance or load end abnormality.

[0118] The arithmetic mean of the multi-dimensional compensation data set of each reference current transformer is calculated to obtain the multi-dimensional compensation data of the disturbed current transformer.

[0119] The initial multi-dimensional data of the disturbed current transformer is updated using exponential smoothing based on the multi-dimensional compensation data of the disturbed current transformer.

[0120] In specific embodiments, for each feature dimension d (current effective value, total harmonic distortion rate, current phase angle and power factor), the specific execution mode of exponential smoothing is as follows: , wherein, is the original value of the disturbed current transformer in the d dimension, is the multi-dimensional compensation data of the disturbed current transformer, is the updated smoothed value of the disturbed current transformer, is the smoothing factor, and d is the feature dimension.

[0121] Each disturbed current transformer is traversed to perform disturbed data update of each disturbed current transformer.

[0122] The fault source current transformer fault diagnosis module is used to record the corresponding preliminary abnormality current transformer as the source fault current transformer when the state execution branch is fault diagnosis, and to perform fault diagnosis of each fault source current transformer based on the machine learning model.

[0123] Further, the fault diagnosis of each fault source current transformer is performed based on the machine learning model, and the specific process is as follows:

[0124] The preliminary abnormality current transformer judged as the source fault label, i.e. the source fault current transformer, triggers the gating mechanism to start the fault diagnosis sub-network inside the model.

[0125] The input feature vectors of each source fault current transformer are multiplexed to perform fault mode recognition tasks based on the fault diagnosis sub-network.

[0126] The output includes fault diagnosis results, including the specific fault type and fault level (Level 1, Level 2, Level 3, etc.).

[0127] In specific embodiments, specific fault types can be divided into several subcategories with identifiable electrical behavior patterns, including the following:

[0128] Inter-turn short-circuit faults refer to the damage to the inter-turn insulation between the primary or secondary windings of a current transformer, resulting in a short circuit in some windings. The symptoms include a persistently low output current, distortion of the secondary current, increased harmonic distortion, and possibly excessively rapid temperature rise. It is a typical structural fault.

[0129] Open-circuit faults typically occur on the secondary side due to open circuits or loose wiring, causing intermittent circuitry or complete signal interruption. Symptoms include interrupted current signals, increased secondary-side voltage, and in severe cases, the risk of arcing. Furthermore, they can significantly impact the stability of the entire data acquisition chain.

[0130] Core saturation faults occur when a transformer core enters the magnetic saturation region due to high current surges, unbalanced operating conditions, or design flaws. These faults often result in output waveform clipping and enhanced odd harmonics, particularly noticeable in scenarios with sudden load changes or frequent system disturbances. In severe cases, they can affect the power factor and metering accuracy.

[0131] Insulation deterioration or breakdown faults are caused by material aging due to long-term operation of the instrument transformer or harsh environment. They are characterized by intermittent abnormal output, unstable signal, and abnormal leakage current or insulation detection alarms that are prone to occur under humid, temperature rise or load change conditions.

[0132] Abnormal transformation ratio or accuracy drift faults are commonly seen in current transformers after long-term operation, such as accuracy degradation, winding loosening, and mechanical fatigue. They manifest as a significant output offset trend, increased deviation from similar equipment, and inability to accurately reflect the true primary side signal even under rated load.

[0133] like Figure 2 The flowchart shown illustrates a machine learning-based fault diagnosis method for integrated current transformers. This method provides a machine learning-based fault diagnosis approach for integrated current transformers, which includes:

[0134] The instantaneous current waveform data of each current transformer is collected, the abnormal offset dataset of each current transformer is calculated, and the abnormality preliminary judgment of each current transformer is performed. The current transformers whose abnormality preliminary judgment result is suspected to be abnormal are recorded as the abnormal current transformers.

[0135] The feature vectors of the initial abnormal current transformers are constructed, and the label probability dataset of the initial abnormal current transformers is output based on the machine learning model, the state labels of the initial abnormal current transformers are analyzed, and the state execution branches of the initial abnormal current transformers are determined.

[0136] When the state execution branch is the disturbed data update, the corresponding initial abnormal current transformer is recorded as a disturbed current transformer, a reference current transformer set is determined, and the disturbed data update of each disturbed current transformer is performed.

[0137] When the state execution branch is the fault diagnosis, the corresponding initial abnormal current transformer is recorded as a source fault current transformer, and the fault diagnosis of each fault source current transformer is performed based on the machine learning model.

[0138] The above embodiments can be implemented by software, hardware (such as a circuit), firmware or any combination thereof, in whole or in part. When implemented by software, the above embodiments can be implemented in the form of a computer program product, in whole or in part. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the flow or function according to the embodiments of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0139] It should be understood that the term "and / or" herein is only a description of the association relationship between the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone, where A and B can be singular or plural. In addition, the character " / " herein generally represents that the associated objects before and after it are in an "or" relationship, but it can also represent an "and / or" relationship, which can be understood according to the context before and after it.

[0140] In the present application, "at least one" means one or more, and "multiple" means two or more. "At least one of the following" or the like means any combination of the items, including a single item or any combination of multiple items. For example, at least one of a, b, or c can mean a, b, c, a-b, a-c, b-c, or a-b-c, where a, b, and c can be single or multiple.

[0141] It should be understood that the size of the sequence number of the above-mentioned processes does not mean the order of execution in various embodiments of the present application. The execution order of the processes should be determined by their functions and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0142] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0143] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the devices, apparatuses and units described above can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0144] In addition, the functional units in various embodiments of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0145] If the functions are realized in the form of software functional units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art, or parts of the technical solutions can be embodied in the form of software products. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0146] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A fault diagnosis system for integrated current transformers based on machine learning, characterized in that, The system includes: The current transformer anomaly preliminary judgment module is used to collect the instantaneous current waveform data of each current transformer, calculate the abnormal offset dataset of each current transformer, and then perform the anomaly preliminary judgment of each current transformer. The current transformer with the anomaly preliminary judgment result as suspected anomaly is recorded as the preliminarily judged abnormal current transformer. The module for judging the status label of the initial abnormal current transformer is used to construct the feature vector of each initial abnormal current transformer, and use this as input to output the label probability dataset of the initial abnormal current transformer based on the machine learning model. It analyzes the status label of each initial abnormal current transformer to determine the status execution branch of each initial abnormal current transformer. The disturbed current transformer disturbed data update module is used to record the corresponding initially judged abnormal current transformer as a disturbed current transformer when the state execution branch is disturbed data update, determine the reference current transformer set, and then perform disturbed data update for each disturbed current transformer. The fault source current transformer fault diagnosis module is used to record the corresponding initially judged abnormal current transformer as the source fault current transformer when the state execution branch is fault diagnosis, and to perform fault diagnosis of each fault source current transformer based on the machine learning model.

2. The machine learning-based fault diagnosis system for coupled current transformers according to claim 1, characterized in that, The instantaneous current waveform data of each current transformer is collected, and the abnormal offset dataset of each current transformer is calculated. The specific calculation process is as follows: Set a sampling period and perform high-frequency sampling on the secondary side of each current transformer to obtain the instantaneous current waveform data of each current transformer; The effective value of the current of each current transformer is calculated based on the instantaneous current waveform data of each current transformer. Based on the instantaneous current waveform data of each current transformer, frequency domain transformation and fundamental frequency extraction are performed simultaneously to calculate the total harmonic distortion rate of each current transformer. Obtain the historical datasets of each current transformer and calculate the historical characterization data of each current transformer, including the historical effective value mean, historical effective value standard deviation, historical average total harmonic distortion rate, and historical total harmonic distortion rate standard deviation. The abnormal offset dataset of each current transformer is calculated based on the effective current value of each current transformer, the total harmonic distortion rate of each current transformer, and the historical characterization data of each current transformer. The abnormal offset dataset for each current transformer includes the abnormal offset of the RMS current value and the abnormal offset of the total harmonic distortion rate for each current transformer.

3. The machine learning-based fault diagnosis system for coupled current transformers according to claim 2, characterized in that, The specific process for performing the initial fault detection of each current transformer is as follows: The abnormal density of each current transformer within a preset historical statistical period is calculated and recorded as the historical abnormal density of each current transformer. The deviation threshold correction coefficient of each current transformer is extracted based on the historical abnormal density of each current transformer. Based on the preset current effective value abnormal offset threshold and total harmonic distortion rate abnormal offset threshold, the deviation threshold correction coefficient of each current transformer is combined to obtain the current effective value abnormal offset correction threshold and total harmonic distortion rate abnormal offset correction threshold of each current transformer. If the abnormal deviation of the effective current value of a current transformer exceeds the abnormal deviation correction threshold of the effective current value of the current transformer or the abnormal deviation of the total harmonic distortion rate exceeds the abnormal deviation correction threshold of the total harmonic distortion rate of the current transformer, then the initial abnormality judgment result of the current transformer is a suspected abnormality, and the current transformer is recorded as an initially judged abnormal current transformer. If the abnormal deviation of the effective current value of a current transformer does not exceed the abnormal deviation correction threshold of the effective current value of the current transformer and the abnormal deviation of the total harmonic distortion rate does not exceed the abnormal deviation correction threshold of the total harmonic distortion rate of the current transformer, then the initial abnormal judgment result of the current transformer is normal, and the current transformer is recorded as the initially judged normal current transformer.

4. The machine learning-based fault diagnosis system for coupled current transformers according to claim 1, characterized in that, The specific process for constructing the feature vectors of each initially identified abnormal current transformer is as follows: Obtain the numbers of each initially identified abnormal current transformer; Extract the first abnormal timestamp of each initially judged abnormal current transformer, and set the earliest abnormal judgment coefficient of the earliest initially judged abnormal current transformer to 1, and set the earliest abnormal judgment coefficient of the other initially judged abnormal current transformers to 0. Based on the first abnormal timestamp of each initially judged abnormal current transformer, the electromagnetic interference intensity of the previous moment corresponding to the first abnormal timestamp of each initially judged abnormal current transformer is extracted and denoted as the electromagnetic interference intensity of each initially judged abnormal current transformer. Collect bus voltage and calculate the current-bus voltage correlation coefficient of each initially identified abnormal current transformer; Extract the abnormal offset dataset of each initially identified abnormal current transformer from the abnormal offset dataset of each current transformer; The number of each initially judged abnormal current transformer, the earliest abnormal judgment coefficient of each initially judged abnormal current transformer, the electromagnetic interference intensity of each initially judged abnormal current transformer, the current-bus voltage correlation coefficient of each initially judged abnormal current transformer, and the abnormal offset dataset of each initially judged abnormal current transformer are combined as the state label analysis feature vector of each initially judged abnormal current transformer. The instantaneous waveform signal characteristics, environmental state context characteristics, and historical anomaly statistical characteristics of each initially identified abnormal current transformer are collected and analyzed, and combined as supplementary feature vectors for each initially identified abnormal current transformer. The state label analysis feature vector of each initially identified abnormal current transformer and the supplementary feature vector of each initially identified abnormal current transformer are combined as the feature vector of each initially identified abnormal current transformer.

5. The machine learning-based fault diagnosis system for coupled current transformers according to claim 4, characterized in that, The specific output process for the tag probability dataset of the initially identified abnormal current transformer is as follows: The feature vectors of each initially identified abnormal current transformer are input into a pre-trained machine learning model. The machine learning model adopts a multi-task MLP gating model and outputs the label probability dataset of each initially identified abnormal current transformer through the main discrimination channel. The tag probability dataset for each initially identified abnormal current transformer includes the fault source probability, disturbance probability, and normal probability of each initially identified abnormal current transformer.

6. The machine learning-based fault diagnosis system for coupled current transformers according to claim 1, characterized in that, The process of analyzing the status labels of each initially identified abnormal current transformer to determine the status execution branch for each initially identified abnormal current transformer is as follows: Based on the label probability dataset of each initially judged abnormal current transformer, the probability type with the highest probability is recorded as the state label of the corresponding initially judged abnormal current transformer. The status labels include source fault labels, disturbed labels, and normal labels; Based on the status labels, each initially judged abnormal current transformer is divided into current transformers with various sources of fault, disturbed current transformers, and normal current transformers. If a certain initially identified abnormal current transformer is a disturbed current transformer, then the state execution branch of the initially identified abnormal current transformer is recorded as disturbed data update. If a certain initially identified abnormal current transformer is a source fault current transformer, then the state execution branch of the initially identified abnormal current transformer is recorded as source fault diagnosis. If a current transformer initially identified as abnormal is found to be a normal current transformer, then the execution branch for the state of the initially identified abnormal current transformer is recorded as "continue sampling and detection".

7. The machine learning-based fault diagnosis system for coupled current transformers according to claim 1, characterized in that, The specific process for determining the set of reference current transformers is as follows: For a disturbed current transformer, if there are no initially identified normal current transformers and normal current transformers, a prompt message is generated and the disturbed data is not updated. If there is a pre-judged normal current transformer or a normal current transformer, the pre-judged normal current transformer or the normal current transformer shall be collectively referred to as each basic reference current transformer. Based on each basic reference current transformer, calculate the similarity between each basic reference and the disturbed current transformer. Based on the abnormal offset dataset of each current transformer, the abnormal offset dataset of each basic reference current transformer is extracted, including the abnormal offset of the current effective value of each basic reference current transformer and the abnormal offset of the total harmonic distortion rate of each basic reference current transformer. The similarity between each base reference-disturbed current transformer and the similarity between the reference base reference-disturbed current transformer are compared. The abnormal deviation of the reference current effective value and the abnormal deviation of the reference total harmonic distortion rate are compared with the abnormal deviation of the current effective value and the abnormal deviation of the total harmonic distortion rate of each base reference current transformer, respectively. The comparison results are then coupled with a weighting factor to obtain the similarity substitution score of each base reference current transformer. Extract the preset similarity substitution scoring threshold from the database; If there is no similar substitute rating greater than or equal to the similar substitute rating threshold, a warning message will be generated and the disturbed data will not be updated. If there is a base reference current transformer with a similar substitution score greater than or equal to the similar substitution score threshold, the corresponding base reference current transformer is recorded as the reference current transformer. By iterating through each reference current transformer, a set of reference current transformers is obtained. By iterating through each disturbed current transformer, the set of reference current transformers corresponding to each disturbed current transformer is obtained.

8. The machine learning-based fault diagnosis system for coupled current transformers according to claim 7, characterized in that, The specific analysis process for updating the disturbance data of each disturbed current transformer is as follows: For a disturbed current transformer, obtain the set of reference current transformers corresponding to the disturbed current transformer, and then obtain the multi-dimensional compensation data set of each reference current transformer. The arithmetic mean of the multidimensional compensation data set of each reference current transformer is calculated to obtain the multidimensional compensation data of the disturbed current transformer. The initial multidimensional data of the disturbed current transformer is smoothed exponentially, and the disturbed data is updated based on the multidimensional compensation data of the disturbed current transformer. Iterate through each disturbed current transformer and update the disturbed data for each disturbed current transformer.

9. The fault diagnosis system for coupled current transformers based on machine learning according to claim 1, characterized in that, The specific process of performing fault diagnosis on each fault-source current transformer based on a machine learning model is as follows: The system acquires the initially identified abnormal current transformers that are identified as source faults, i.e., source fault current transformers, and triggers the gating mechanism to start the fault diagnosis sub-network inside the model. The input feature vectors of fault current transformers from various sources are reused to perform fault mode recognition tasks based on the fault diagnosis subnetwork. The output includes fault diagnosis results, including the specific fault type and fault level.

10. A machine learning-based fault diagnosis method for integrated current transformers, applied to the machine learning-based fault diagnosis system for integrated current transformers as described in any one of claims 1-9, characterized in that, The method includes the following steps: Collect instantaneous current waveform data of each current transformer, calculate the abnormal offset dataset of each current transformer, and then perform an initial anomaly judgment for each current transformer. Record the current transformers whose initial anomaly judgment result is suspected to be abnormal as the initial anomaly current transformers. Construct feature vectors for each initially identified abnormal current transformer, and use these as input to output the label probability dataset of the initially identified abnormal current transformer based on the machine learning model. Analyze the state labels of each initially identified abnormal current transformer to determine the state execution branch of each initially identified abnormal current transformer. When the state execution branch is for disturbed data update, the corresponding initially judged abnormal current transformer is recorded as disturbed current transformer, the reference current transformer set is determined, and the disturbed data update of each disturbed current transformer is performed. When the state execution branch is fault diagnosis, the corresponding initially judged abnormal current transformer is recorded as the source fault current transformer, and the fault diagnosis of each fault source current transformer is performed based on the machine learning model.

Images