Precision robust transient fault component extraction method for ultra-high frequency sampling

Abstract

The application relates to the technical field of power system relay protection, and discloses a precision-robust transient fault component extraction method for super-high-frequency sampling, which comprises the following steps: obtaining healthy historical information before a fault and saving the information in a low-speed data area; obtaining real-time frequency estimation values and cycle estimation values corresponding to a fault starting moment; according to the cycle estimation values, backstepping and superimposing time alignment compensation values at any high-frequency sampling moment after the fault to construct a reference moment before the fault corresponding to the high-frequency sampling moment; according to the steady-state historical information before the fault, adaptively selecting a reconstruction mode and reconstructing a reference waveform before the fault at the reference moment before the fault; and subtracting high-speed original sample values at the high-frequency sampling moment after the fault from the reference waveform before the fault at the reference moment before the fault to obtain a transient fault component. The application improves the accuracy and robustness of the transient fault component extraction under the conditions of frequency deviation, harmonics and waveform distortion.

Description

Technical Field

[0001] This invention relates to the field of power system relay protection technology, and more specifically, to a method for extracting transient fault components with high accuracy and robustness for ultra-high frequency sampling. Background Technology

[0002] In recent years, the extensive integration of power electronic equipment into new energy transmission systems, flexible AC transmission systems, and flexible low-frequency transmission systems has accelerated the development of voltage and current transient processes after faults, significantly shortening the action time of current limiting, blocking, and mode switching in control systems. Traditional protection methods based on power frequency quantities rely on long data windows, making it difficult to simultaneously meet the speed and reliability requirements of ultra-high-speed protection in new power systems.

[0003] The fault component method constructs the changes in electrical quantities after a fault relative to the corresponding time before the fault, thus forming a relatively simple circuit relationship. It is suitable for scenarios such as direction determination, differential fault determination, fault phase selection, and auxiliary ranging. For ultra-high-speed protection, the timeliness and accuracy of fault component extraction directly determine the performance of subsequent protection criteria.

[0004] Existing methods for constructing pre-fault reference values ​​can be broadly categorized into three types. The first type stores pre-fault integer or multi-cycle high-frequency raw data and retrieves the reference value after the fault by backtracking at fixed sampling points. This method is structurally intuitive, but suffers from significant storage costs and data transfer burdens under ultra-high frequency sampling conditions. The second type parameterizes the pre-fault steady-state waveform, directly reconstructing the pre-fault reference value using parameters such as frequency, amplitude, and phase. This method has lower storage overhead, but is prone to model mismatch errors when frequency offset, harmonics, and waveform distortion are present. The third type further incorporates historical sample values ​​to correct residuals based on parameterized reconstruction, improving accuracy. However, existing publicly available solutions lack a clear and complete processing chain for buffer organization, time alignment, and engineering implementation in protection devices.

[0005] Furthermore, frequency offset and frequency fluctuation phenomena are more prevalent in new power systems. If the fault component is still extracted using a fixed integer frequency and fixed sampling point backtracking method, the high-frequency sampling time after the fault is prone to misalignment with the corresponding reference time before the fault, thus introducing additional deviations into the fault component.

[0006] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0007] To address the problems in related technologies, this invention proposes a method for extracting transient fault components with high accuracy for ultra-high frequency sampling, thereby overcoming the aforementioned technical problems in existing related technologies.

[0008] Therefore, the specific technical solution adopted by the present invention is as follows:

[0009] A method for extracting transient fault components with high accuracy for ultra-high frequency sampling includes:

[0010] S1. Obtain the original high-frequency sample sequence of multiple channels, save the original high-frequency sample sequence of multiple channels in the preset high-speed data area and low-speed data area, and downsample or resample the original high-frequency sample sequence of multiple channels to obtain the health history information before the fault and save it in the low-speed data area; the health history information before the fault includes the steady-state history information before the fault, parameterized steady-state information, data quality indicators of each channel and channel status flags;

[0011] S2. Perform online frequency measurement based on the health history information before the fault to obtain the real-time frequency estimate and period estimate corresponding to the fault start time;

[0012] S3. Based on the period estimate, backtrack any high-frequency sampling time after the fault and superimpose the time alignment compensation to construct the reference time before the fault corresponding to the high-frequency sampling time.

[0013] S4. Based on the steady-state history information before the fault, adaptively select the reconstruction mode and reconstruct the reference waveform before the fault at the reference time before the fault.

[0014] S5. Subtract the high-speed original sample value at the high-frequency sampling time after the fault from the reference waveform at the reference time before the fault to obtain the transient fault component.

[0015] The beneficial effects of this invention are as follows:

[0016] 1. This invention significantly reduces storage occupancy and data transfer pressure in ultra-high frequency sampling scenarios by setting up high-speed and low-speed data areas and storing high-frequency raw data only in the fault neighborhood. By introducing online frequency measurement, pre-fault reference time construction, and multi-mode reference waveform reconstruction mechanisms, it improves the accuracy and robustness of transient fault component extraction under frequency offset, harmonics, and waveform distortion conditions. By separating high-speed cache management from reference waveform reconstruction, it is easy to implement on a heterogeneous platform composed of FPGA, DSP, and CPU, and has good engineering applicability.

[0017] 2. This invention reduces reference waveform errors caused by using only fixed-period backtracking or single-parameter reconstruction by combining parameterized reference waveforms and pre-fault residual sequences for correction. Channel status flags and data quality indicators prevent contamination of low-speed parameter estimation and pre-fault reference waveform reconstruction caused by single analog channel disconnections, saturation, distortion, or sampling link anomalies. Through condition-adaptive reconstruction mode selection and residual segmented weighted correction, the robustness of transient fault component extraction under frequency offset, harmonics, noise, and local channel anomaly conditions is improved.

[0018] 3. This invention distinguishes hardware faults, electromagnetic interference, and line defects by tracing the fingerprint of sampled data, and completes the reconstruction mode selection through a fuzzy rough set operating condition comprehensive evaluation mechanism. This makes channel anomaly handling and mode switching no longer dependent on a single amplitude threshold or fixed threshold comparison, thereby improving the accuracy and interpretability of the reference waveform reconstruction before the fault under critical operating conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a method for extracting transient fault components with high accuracy for ultra-high frequency sampling according to an embodiment of the present invention;

[0021] Figure 2 This is a block diagram of a high-precision, robust transient fault component extraction system for ultra-high frequency sampling according to an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of a dual-memory area structure according to an embodiment of the present invention;

[0023] Figure 4 This is a time axis diagram illustrating the extraction of transient fault components by combining the high-speed data area and the low-speed data area according to an embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram of the construction of the reference time before the fault according to an embodiment of the present invention;

[0025] Figure 6 This is a schematic diagram of the reconstruction of the reference waveform and residual correction combination according to an embodiment of the present invention;

[0026] Figure 7 This is a schematic diagram of three reference waveform reconstruction modes according to an embodiment of the present invention;

[0027] Figure 8 This is a schematic diagram of the FPGA and DSP / CPU collaborative deployment structure according to an embodiment of the present invention.

[0028] In the picture:

[0029] 1. Dual storage management module; 2. Online frequency measurement module; 3. Reference time construction module; 4. Reference waveform reconstruction module; 5. Fault component calculation module. Detailed Implementation

[0030] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0031] According to embodiments of the present invention, a method for extracting transient fault components with high accuracy for ultra-high frequency sampling is provided.

[0032] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the method for extracting accurate and robust transient fault components for ultra-high frequency sampling according to an embodiment of the present invention includes:

[0033] S1. Obtain the original high-frequency sample sequence of multiple channels, save the original high-frequency sample sequence of multiple channels in the preset high-speed data area and low-speed data area, and downsample or resample the original high-frequency sample sequence of multiple channels to obtain the health history information before the fault and save it in the low-speed data area; the health history information before the fault includes the steady-state history information before the fault, parameterized steady-state information, data quality indicators of each channel and channel status flags;

[0034] S2. Perform online frequency measurement based on the health history information before the fault to obtain the real-time frequency estimate and period estimate corresponding to the fault start time;

[0035] S3. Based on the period estimate, backtrack any high-frequency sampling time after the fault and superimpose the time alignment compensation to construct the reference time before the fault corresponding to the high-frequency sampling time.

[0036] S4. Based on the steady-state history information before the fault, adaptively select the reconstruction mode and reconstruct the reference waveform before the fault at the reference time before the fault.

[0037] S5. Subtract the high-speed original sample value at the high-frequency sampling time after the fault from the reference waveform at the reference time before the fault to obtain the transient fault component.

[0038] In this optional embodiment, S1 includes:

[0039] S11. A unified sampling clock is used to trigger the analog-to-digital conversion of each channel, and the multi-channel voltage and current analog quantities of the protected object are sampled synchronously at ultra-high frequency to form the multi-channel original high-frequency sample vector and its timestamp corresponding to the current sampling sequence number.

[0040] S12. Save the original high-frequency sample vectors of the multi-channel in the circular buffer of the high-speed data area. The write pointer increments by taking the modulo of the length of the circular buffer after each new sample vector arrives. After the circular buffer is full, the oldest original high-frequency sample vector is overwritten.

[0041] S13. After performing anti-aliasing low-pass filtering on the original high-frequency sample vectors of the multi-channel, extract or resample them according to the preset downsampling ratio to obtain low-speed historical sample values ​​and save them in the low-speed data area.

[0042] S14. Periodically update the parameterized steady-state information of each channel based on the low-speed historical sample values ​​and save it in the low-speed data area. The parameterized steady-state information includes at least one of frequency, amplitude, phase, main harmonic coefficient and DC bias.

[0043] S15. Save the channel status flags and data quality indicators generated for each channel in the high-speed data area and the low-speed data area. The channel status flags are used to characterize whether there is a disconnection, saturation, abnormal sampling link, abrupt distortion, or excessive noise in the channel. Generate a sampling data traceability fingerprint for each set of original high-frequency samples and save it in the low-speed data area. The sampling data traceability fingerprint includes the sampling link number, analog-to-digital conversion saturation flag, clock node deviation, synchronization pulse jitter, environmental interference intensity, correlation between adjacent channels, and historical anomaly type.

[0044] In this optional embodiment, S2 includes:

[0045] S21. Using the main tracking channel as the object, digital phase-locked loop is used to obtain the initial values ​​of the fundamental phase angle and angular frequency;

[0046] S22. Within the sliding health data window, recursive least squares estimation is used to correct the parameters of the fundamental wave model to obtain the fundamental wave amplitude, phase correction amount, and DC bias.

[0047] S23. Update the frequency estimate once every preset update interval, and write the updated frequency estimate, fundamental amplitude, phase correction and DC bias into the parameterized steady-state information of the low-speed data area.

[0048] S24. Based on the change in phase correction at adjacent update times, the base frequency given by the digital phase-locked loop is corrected to obtain the real-time frequency estimate.

[0049] S25. At the moment of fault initiation, determine the corresponding cycle estimate based on the real-time frequency estimate.

[0050] In this optional embodiment, S22 includes:

[0051] S221. Obtain the low-speed historical sample values ​​in the sliding health data window as the data source for recursive least squares estimation. The low-speed historical sample values ​​are the historical sample values ​​before the fault starts, after anti-aliasing filtering, abnormal channel elimination and timestamp correction.

[0052] S222. Preprocess the low-speed historical samples, including removing the initial DC bias value, assigning weights to the samples according to the data quality index, and removing samples marked as unusable by the channel status flag.

[0053] S223. Based on the preprocessed low-speed historical samples, using the fundamental wave model containing cosine, sine and DC bias terms as the framework, and aiming to minimize the weighted squared residual, the parameters of the fundamental wave model are recursively estimated by recursive least squares update with forgetting factor, and the recursively obtained parameters are converted into fundamental wave amplitude, phase correction and DC bias.

[0054] In this optional embodiment, S223 includes:

[0055] S2231. Set the initial covariance matrix and forgetting factor; the initial covariance matrix is ​​the product of a scalar and an identity matrix. The scalar is selected based on the initial parameter uncertainty, the length of the health window, and the noise level. The forgetting factor is selected based on the noise level of the health window before the fault and the required tracking speed.

[0056] S2232. Construct the regression vector and parameter vector; the regression vector consists of the cosine, sine, and constant terms corresponding to the current sample value, and the parameter vector consists of the parameters to be estimated in the fundamental wave model. The phase angles in the cosine and sine terms are obtained by integrating the initial phase angle and angular frequency given by the digital phase-locked loop.

[0057] S2233. Obtain each preprocessed historical sample value within the sliding health data window according to the timestamp, obtain the weight of each sample value by normalization of the sample value quality factor, and remove the sample points marked as unusable by the channel status flag.

[0058] S2234. Construct an objective function with the goal of minimizing the weighted squared residual with forgetting factor, and update the objective function with the reduction of the weighted residual of the objective function as the parameter of the recursive process.

[0059] S2235. Calculate the current recursive gain based on the regression vector, new input sample and its weight corresponding to the current recursive update index, and in combination with the parameter vector and covariance matrix of the previous time step.

[0060] S2236. Update the parameter vector at the current time step based on the current recursive gain, the parameter vector at the previous time step, and the estimation error of the new input sample.

[0061] S2237. Based on the current recursive gain, the covariance matrix of the previous time step, and the current regression vector, update the covariance matrix of the current time step. The updated covariance matrix is ​​used for the next recursive gain calculation and parameter reliability judgment.

[0062] S2238. Calculate the fundamental amplitude and phase correction based on the cosine and sine coefficients in the parameter vector at the current time, and obtain the DC bias based on the DC term in the parameter vector at the current time.

[0063] In this optional embodiment, S4 includes:

[0064] S41. Calculate the frequency offset index, harmonic distortion index, noise level index, data quality index, residual consistency index, and computational load index respectively.

[0065] S42. Based on the calculation results of each index, adaptively select any reconstruction mode; if the frequency offset and harmonic distortion are small and computing resources are scarce, select the direct parameter reconstruction mode; if the parameter estimation link is unavailable but the quality of low-speed historical samples is high, select the sample interpolation recovery mode; if the frequency offset, harmonic distortion or waveform distortion are significant and the residual sequence is continuous, select the hybrid mode of parameter reconstruction plus residual correction.

[0066] S43. When the sample interpolation recovery mode is selected, several low-speed historical sample points are selected near the reference time before the fault, and the interpolation method is used to recover the reference waveform before the fault at the reference time before the fault.

[0067] S44. When the parameter direct reconstruction mode is selected, the reference waveform before the fault at the reference time before the fault is reconstructed based on the parameterized steady-state information in the low-speed data area. The parameterized steady-state information includes at least one of frequency, amplitude, phase, harmonic coefficient and DC bias.

[0068] S45. When the hybrid mode of parameter reconstruction plus residual correction is selected, the reference waveform before the fault is reconstructed according to the parameterized steady-state information, and the residual sequence is constructed according to the difference between the low-speed historical sample and the reference waveform. The residual correction amount is estimated near the reference time before the fault. The reference waveform and the residual correction amount are superimposed to obtain the reference waveform before the fault at the reference time before the fault.

[0069] In this optional embodiment, estimating the residual sequence near the reference time before the fault to obtain the residual correction amount includes:

[0070] Before and after the reference time before the fault, select residual candidate points that meet the requirements of valid channel status flag and data quality index not lower than the preset lower limit to form a residual candidate set.

[0071] The residual candidate points are segmented and labeled according to the residual amplitude, local variance, adjacent residual slope, and source fingerprint of the sampled data.

[0072] For candidate points marked as normal fluctuation residuals, local quadratic least squares fitting with weights determined by time distance and data quality indicators is used, or local spline interpolation is used when the residual continuity meets the preset conditions to obtain the residual correction amount at the reference time before the fault.

[0073] Candidate points marked as distorted residuals and noise-exceeding residuals are reduced in weight or removed from the residual candidate set. For unusable residuals, the previous effective correction amount or zero correction amount is used as the current residual correction amount.

[0074] According to another embodiment of the invention, such as Figure 2 As shown, a robust transient fault component extraction system for ultra-high frequency sampling is also provided, comprising:

[0075] The dual storage management module 1 is used to acquire multi-channel raw high-frequency sample sequences, store the multi-channel raw high-frequency sample sequences in a preset high-speed data area and low-speed data area, downsample or resample the multi-channel raw high-frequency sample sequences to obtain pre-fault health history information and store it in the low-speed data area; the pre-fault health history information includes pre-fault steady-state history information, parameterized steady-state information, data quality indicators of each channel, and channel status flags;

[0076] Online frequency measurement module 2 is used to perform online frequency measurement based on the health history information before the fault, and obtain the real-time frequency estimate and period estimate corresponding to the fault start time;

[0077] Reference time construction module 3 is used to backtrack any high-frequency sampling time after the fault and superimpose the time alignment compensation amount based on the period estimate to construct the reference time before the fault corresponding to the high-frequency sampling time.

[0078] Reference waveform reconstruction module 4 is used to adaptively select the reconstruction mode and reconstruct the reference waveform before the fault at the reference time based on the steady-state history information before the fault.

[0079] The fault component calculation module 5 is used to subtract the high-speed original sample value at the high-frequency sampling time after the fault from the reference waveform at the reference time before the fault to obtain the transient fault component.

[0080] To facilitate understanding of the above technical solutions of the present invention, the following provides a detailed description of the accurate and robust transient fault component extraction for ultra-high frequency sampling in practical applications.

[0081] It needs to be explained that, for example Figure 3 As shown, this invention focuses on a three-phase AC line protection device for flexible low-frequency power transmission or new energy transmission scenarios. The protection device synchronously collects three-phase voltage data. , , and three-phase current , , It has a total of 6 analog inputs. The analog-to-digital conversion unit uses a unified clock for synchronous sampling, with a sampling frequency of 1MHz and a corresponding sampling interval of 1μs. The device adopts a dual-storage area structure: the high-speed data area is used to store the original high-frequency sample values ​​before and after the fault, and the low-speed data area is used to store the steady-state historical sample values ​​and parameter information before the fault.

[0082] In other implementations, the original sampling rate of the high-speed data area can be tens of kHz to several MHz depending on the protected object, the transient frequency band range, and the analog-to-digital conversion hardware capability; the equivalent sampling rate of the low-speed data area is lower than that of the high-speed data area, and steady-state historical information before the fault is preserved by downsampling or resampling.

[0083] Each sampling generates a 6-channel raw high-frequency sample vector with a sampling sequence number and a timestamp. This sample vector serves as the data source for both the high-speed and low-speed data areas.

[0084] The high-speed data area stores the original high-frequency samples 5ms before and 10ms after the fault, therefore each channel needs to store 15,000 high-frequency sampling points. The low-speed data area stores historical information of the most recent two electrical cycles at an equivalent rate of 2kHz, corresponding to a 40ms history window at a 50Hz operating frequency, therefore each channel needs to store 80 low-speed samples. If the traditional scheme directly stores the original data of two cycles at a sampling rate of 1MHz for a long time, each channel needs to store 40,000 sampling points, and the six channels require a total of about 468.8kB of storage. After adopting the dual storage area structure of this invention, each channel only needs to store 15,080 sampling points, and the six channels require a total of about 176.7kB of storage, reducing the storage amount by about 62.3%.

[0085] The high-speed data area stores the original high-frequency samples 5ms before and 10ms after the fault, therefore each channel needs to store 15,000 high-frequency sampling points. The low-speed data area stores historical information of the most recent two electrical cycles at an equivalent rate of 2kHz, corresponding to a 40ms history window at a 50Hz operating frequency, therefore each channel needs to store 80 low-speed samples. If the traditional scheme directly stores the original data of two cycles at a sampling rate of 1MHz for a long time, each channel needs to store 40,000 sampling points, and the six channels require a total of about 468.8kB of storage. After adopting the dual storage area structure of this invention, each channel only needs to store 15,080 sampling points, and the six channels require a total of about 176.7kB of storage, reducing the storage amount by about 62.3%.

[0086] During normal operation, the device continuously performs ultra-high frequency synchronous sampling on 6 analog signals and writes each newly arrived high-frequency sample value into the circular buffer of the high-speed data area. At the same time, the low-speed data area extracts or resamples the original high-frequency samples according to the preset downsampling ratio, retains the main information of the steady-state historical waveform before the fault, and periodically updates the frequency, amplitude, phase, main harmonic coefficients and DC bias of each channel.

[0087] Specifically, after the unified sampling clock triggers the 6-channel analog-to-digital conversion, the current sampling sequence number is generated. The corresponding original high-frequency sample vector and its timestamp The high-speed data area uses a circular buffer structure to store the original high-frequency sample vector, with write pointers... After each new sample vector arrives, the buffer length is used for... Increasing modulus, i.e. And after the buffer is full, it overwrites the oldest original high-frequency sample.

[0088] The low-speed data area is downsampled according to the preset downsampling factor. The original high-frequency samples are decimated or resampled. During processing, the original high-frequency samples are first subjected to anti-aliasing low-pass filtering, and then... A low-speed sample value is generated from each high-frequency sampling point; when historical sample values ​​for non-integer low-speed sampling times are needed, resampling is performed using cubic Lagrange interpolation or local spline interpolation. Specifically, if... If there are at least two valid low-velocity historical samples on both sides and the time interval between the samples is approximately uniform, then cubic Lagrange interpolation is used; if the samples have non-uniform intervals, local slope abrupt changes, or significant differences in data quality between adjacent samples, then local spline interpolation is used, and... As sample point weights; if the low-speed samples come from equally spaced resampling after anti-aliasing low-pass filtering and the frequency band boundaries are known, then band-limited interpolation is used. The resulting interpolation result is used as the sample interpolation recovery mode. .

[0089] The main information includes one or more of the following: low-speed historical samples, sample timestamps, frequency estimates, period estimates, amplitude, phase, major harmonic coefficients, DC bias, local fitting parameters, residual sequences, and data quality indicators. The low-speed data region uses a sliding healthy data window to periodically update parameterized steady-state information: after each preset update interval, the online frequency measurement module obtains a frequency estimate based on the main tracking channel; under the constraint of this frequency estimate, recursive least squares estimation or discrete Fourier estimation is performed on each channel to obtain the fundamental amplitude, phase, major harmonic coefficients, and DC bias of each channel.

[0090] Specifically, the device for the first Each analog channel generates a channel status flag. and data quality indicators Channel status flags are used to characterize whether the channel has a disconnection, saturation, sampling link anomaly, abrupt distortion, or excessive noise. The high-speed data area stores the original high-frequency sample vector. Simultaneously, the corresponding channel status flags are saved for post-fault waveform recording and diagnostic analysis; during downsampling, online frequency measurement, and parameterized steady-state information updates in the low-speed data area, according to... and Abnormal channels are removed, downweighted, or frozen during updates to prevent a single abnormal channel from contaminating the reference waveform prior to the fault. Furthermore, the channel status monitoring module generates a sampling data traceability fingerprint for each set of original high-frequency samples. Sampling data traceability fingerprint This includes sampling link number, analog-to-digital conversion saturation flag, clock node deviation, synchronization pulse jitter, environmental interference intensity, adjacent channel correlation, and historical anomaly types. Based on... Anomalies are categorized into three types: hardware failure, electromagnetic interference, and line defects. For hardware failure anomalies, the corresponding channel is directly removed and a backup channel is switched. For electromagnetic interference anomalies, the weight of the sample points in that time period is reduced and the fitting window is shortened. For line defect anomalies, the fault-related mutation information is retained, but its contribution to the update of frequency and steady-state parameters is frozen.

[0091] When the primary tracking channel is in an abnormal state, the online frequency measurement module switches to a backup effective channel, or performs fusion frequency measurement based on the phase continuity and data quality indicators of multiple effective channels. The length of the sliding health data window is adaptively adjusted according to the number of effective channels, channel status stability, noise level, and frequency fluctuation. When the channel status is stable and the noise is low, a longer health window can be used to improve the stability of parameter estimation. When the frequency fluctuates rapidly or the channel quality deteriorates, the health window can be shortened to improve tracking real-time performance.

[0092] II. The fault initiation criterion consists of a sudden current change, wavefront energy, differential amplitude, or a combination thereof. When any initiation quantity exceeds the threshold, the system records the fault initiation time. Immediately freeze the short-term window data before the fault in the high-speed data area, and continue to store the short-term window data after the fault. Since the original high-frequency samples are only retained in the fault neighborhood, the wavefront and initial transient details can be fully preserved, while avoiding the resource waste caused by storing whole-cycle historical data for a long time at the ultra-high frequency sampling rate.

[0093] The current mutation is the difference between the current high-frequency current sample value and its reference value before the fault or the sample value one cycle ago. The wavefront energy is the sum of the squares of the high-frequency difference components or wavelet high-frequency components within a short time window. The differential amplitude is the absolute value of the difference between adjacent original high-frequency samples or its multi-point smoothing result. The combination includes any initiation quantity exceeding the corresponding threshold, multiple normalized initiation quantities exceeding the comprehensive threshold after weighted summation, or multiple initiation quantities satisfying a preset voting rule. The threshold is determined based on the noise standard deviation, adaptive coefficient, and preset lower limit of the pre-fault health data.

[0094] Short-term window data before failure refers to Previously preset time windows contain multiple channels of raw high-frequency samples, sampling sequence numbers, and timestamps; short-time window data after a fault refers to data from... The original high-frequency samples, sampling sequence number, and timestamp of multiple channels within a preset time window extending backward are included. In this invention, the wavefront refers to the steep leading edge of voltage or current that first appears in the original high-frequency sample sequence after a fault occurs; the initial transient details include the wavefront arrival time, short-time high-frequency oscillation, differential peak value, traveling wave reflection-related fluctuations, and the initial changes of aperiodic components.

[0095] Full-cycle historical data refers to the original high-frequency sample values ​​obtained by continuously saving one or more electrical cycles prior to a fault at an ultra-high-frequency sampling rate. This invention does not store complete high-frequency full-cycle historical data for a long time. Instead, it reconstructs the reference waveform before the fault using low-speed historical samples and parameterized steady-state information, thereby reducing storage usage while preserving wavefront and initial transient details.

[0096] 3. Perform online frequency measurement using pre-fault health data in the low-speed data area. Using phase A voltage as the main tracking channel, the frequency estimate is updated every 0.5ms using a combination of digital phase-locked loop and recursive estimation.

[0097] The method combining digital phase-locked loop (PLL) and recursive estimation involves: first, obtaining initial values ​​of the fundamental phase angle and angular frequency using PLL; then, correcting the fundamental model using recursive least squares estimation within a sliding health data window. The frequency estimate is updated every 0.5 ms, and the updated frequency estimate, fundamental amplitude, phase correction, and DC bias are written into the parameterized steady-state information in the low-speed data region. This online frequency measurement result and parameterized steady-state information are used for subsequent construction of the pre-fault reference time and reconstruction of the pre-fault reference waveform, respectively, and their functional boundaries are independent of the synchronous control of the grid-connected converter.

[0098] Specifically, the data source for recursive least squares estimation is low-speed historical samples within the sliding health data window before fault initiation, after anti-aliasing filtering, abnormal channel removal, and timestamp correction; preprocessing includes removing initial DC bias values, and... Assign weights to samples and remove samples Samples marked as unavailable. The fundamental wave model is written as... ,in The parameters to be estimated are obtained by integrating the initial phase angle and angular frequency given by the digital phase-locked loop. , and The recursive estimation aims to minimize the weighted squared residuals, employing recursive least squares updates with a forgetting factor. The recursively obtained parameters are then converted into fundamental amplitude, phase correction, and DC bias. The forgetting factor... Take values ​​from 0.96 to 0.995, initial covariance The system uses an identity matrix ranging from 100 to 10000 times its original size, with an update interval of 0.5 ms. When the root mean square of the residual exceeds a preset multiple of the standard deviation of the noise within the healthy window twice consecutively, or when the frequency estimate difference between the backup channel and the main tracking channel exceeds 0.05 Hz, the current parameter update is frozen, and the previous valid frequency estimate is adopted. These settings allow the protection device to record and verify the frequency estimation process, data sources, preprocessing, loss function, optimization method, and key parameters.

[0099] In recursive least squares frequency estimation, the time decay, parameter gain, and covariance updates of historical samples are recursively derived according to the following relationships. Let the regression vector... Depend on , Together with the constant term 1, it forms the parameter vector. Depend on , and The weighted objective function with a forgetting factor is:

[0100]

[0101] In the formula, Indicates the first One historical sample value; Indicates the first The weights of each historical sample value are determined by data quality indicators. Normalization yields, Samples marked as unavailable are not included in the summation; Indicates the first The regression vector corresponding to each historical sample value is derived from... , Composed of the constant term 1; the current regression vector in equations (3) to (5) By current update number The value, the transpose of the regression vector and Multiplying gives the first The model estimation term corresponding to each historical sample value; For the reason , and The parameter vector formed; Indicates the parameter estimate; Indicates the first The objective function of the weighted residual sum of squares with forgetting factor at the next update is updated by recursively updating the objective function with the reduction of the weighted residual of the objective function as the parameter, and the current parameter estimate is obtained by equation (3) and equation (4); the fitting quality can be monitored based on the objective function value or its residual term. Indicates the first The historical sample value at the th The time decay weight in the next update The earlier The larger the value, the smaller the weight. The initial covariance and recursive updates are shown in equations (2) to (5):

[0102]

[0103]

[0104]

[0105]

[0106] In equations (1) to (5), The current recursive update sequence number, For the historical sample value sequence number within the health data window; For the first The next update includes new low-speed input samples. For the first The weights of the newly updated low-speed input samples; This is the regression vector corresponding to the current sample value. It is obtained by integrating the initial phase angle and angular frequency given by the digital phase-locked loop; For the recursive gain. In equation (2), The initial covariance matrix, The identity matrix has the same dimension as the parameter vector. Represents scalar With unit array The initial diagonal covariance matrix formed by multiplication; The value ranges from 100 to 10000. The specific value is selected based on the initial parameter uncertainty, the length of the health window, and the noise level. A larger value is taken when the initial uncertainty is large, and a smaller value is taken when the initial parameters are relatively reliable. The values ​​are taken between 0.96 and 0.995 based on the noise level of the health window before the fault and the required tracking speed.

[0107] From equations (1) to (5), it can be seen that As the initial uncertainty, it enters the first recursive gain. The calculation, A larger δ indicates greater uncertainty about the initial parameters, resulting in a larger initial gain and faster convergence; a smaller δ indicates a smaller initial gain and a smoother estimation change. Equation (5) yields... For the first The updated covariance matrix is ​​used to calculate the new gain in the next recursion and to characterize the uncertainty of the current parameter estimate; when As the value gradually decreases with recursion, it indicates that the parameter estimate tends to stabilize. Further updates are needed. It appears simultaneously in the objective function weight and covariance update formulas, and is used to gradually reduce the influence of old sample values ​​on the current parameters.

[0108] After each recursive update The parameterized steady-state information in the low-speed data region is generated through parameter transformation. Specifically, the fundamental amplitude... and phase correction According to formula (6) , DC bias is obtained. Depend on The DC term is obtained and written to the low-speed data area; where, This represents the phase deviation of the recursive model relative to the digital phase-locked loop phase angle.

[0109]

[0110] In equation (6), and These are respectively used as the fundamental amplitude in equation (11) and fundamental phase The updated value; As the DC bias in equation (11) For harmonic terms and The same recursive framework is used to update the data under the corresponding harmonic regression vector, or the discrete Fourier estimation results are written into the low-speed data area after consistency verification.

[0111] When the fundamental frequency given by the digital phase-locked loop is At that time, the phase correction change between adjacent update times is used to correct the frequency estimate, and the real-time frequency estimate is obtained according to equation (7):

[0112]

[0113] In equation (7), This is a phase expansion operator used to map the phase correction difference between adjacent update times to continuous phase branches, avoiding phase skipping. A jump occurs at this time; For the first The base frequency provided by the digital phase-locked loop at the time of the next update is in Hz. For the first The real-time frequency estimate after phase correction at the next update, in Hz; Indicates the first The update time is measured in seconds. This represents the phase correction amount obtained from the recursive least squares estimation, in rad.

[0114] Equation (7) yields At the moment of fault initiation Values Enter equation (8) for calculation ; Enter formula (9) construction ; , and Entering (11) formation ; Then, going through equations (12) to (13), we get... Finally, it is calculated by equation (14). The recursive process is synchronously preserved. , and residual index, among which Used for this parameter calibration. Used for the next gain calculation and parameter reliability assessment; when If the residual index is abnormal, freeze the parameter update or revert to the sample interpolation recovery mode.

[0115] The estimated period corresponding to the fault initiation time is determined according to equation (8):

[0116]

[0117] Measured during fault start-up Therefore, we get Compared to directly using the fixed period corresponding to the rated frequency, equation (8) can reflect the periodic changes caused by the frequency deviation of the low-frequency system.

[0118] In equation (8), Indicates the time of fault initiation The real-time frequency estimate is obtained from the health data prior to the failure. This represents the estimated period corresponding to the real-time frequency.

[0119] IV. Construct a reference time before the fault for any high-frequency sampling time after the fault, such as Figure 4 and Figure 5 As shown, the sample value after the fault is backtracked by one real-time period according to equation (9), and the time alignment compensation is superimposed. .

[0120] Specifically, for any high-frequency sampling moment after the fault starts First, estimate the real-time cycle value corresponding to the fault start time. The sampling time is pushed back one electrical cycle, and then the time alignment compensation is added. To obtain the reference time before the failure .

[0121]

[0122] For example, at a certain high-frequency sampling time after the fault If obtained through error correction by back-substituting health data Then the reference time before the fault Since this moment typically does not coincide with integer sampling points in the low-speed data region, reference waveform reconstruction is required.

[0123] In equation (9), Indicates the high-frequency sampling time after the fault. Indicates and The corresponding reference time before the fault, This indicates the time alignment compensation amount. The compensation amount is determined based on the back-substitution error of the pre-fault health data: several known pre-fault sample values ​​are back-substituted and reconstructed using candidate compensation amounts, and the compensation amount that minimizes the back-substitution error is selected as the current compensation amount. It can be iteratively updated during operation.

[0124] V. For example Figure 7 As shown, the reference waveform before the fault is reconstructed based on the steady-state historical information in the low-speed data region. Figure 7 The selection conditions, input information, and output relationships for three reference waveform reconstruction modes are shown. To balance implementation complexity, real-time performance, and accuracy, this invention simultaneously supports three modes: sample interpolation recovery mode, direct parameter reconstruction mode, and a hybrid mode combining parameter reconstruction and residual correction.

[0125] The reference waveform reconstruction module first calculates frequency offset, harmonic distortion, noise level, data quality, residual consistency, and computational load indicators using the operating condition identification module. If the frequency offset and harmonic distortion are small and computational resources are limited, the direct parameter reconstruction mode is prioritized. If the parameter estimation link is unavailable but the quality of low-speed historical samples is high, the sample interpolation recovery mode is selected. If the frequency offset, harmonic distortion, or waveform distortion is significant and the residual sequence is continuous, a hybrid mode combining parameter reconstruction and residual correction is selected. To avoid frequent switching caused by relying solely on fixed thresholds under critical operating conditions, the operating condition identification module also constructs a fuzzy rough set operating condition comprehensive evaluation table: first, frequency offset, harmonic distortion, noise level, data quality, residual consistency, and computational load are mapped to three categories of fuzzy semantic quantities: low, medium, and high; then, attribute dependencies are calculated based on the decision table formed by the historical health window and the low-speed data before the current fault, and redundant indicators with low contribution to the mode decision are reduced; finally, the confidence scores of the three reconstruction modes are output based on the reduced indicator combinations. If the difference between the highest confidence level and the second highest confidence level is less than the preset margin, the hybrid mode of parameter reconstruction plus residual correction is selected first, and the rate of change of residual weight is reduced to maintain the continuity of mode switching.

[0126] Specifically, the domain of the comprehensive evaluation table for operating conditions The evaluation sample consists of historical health windows and low-speed data prior to the current fault, and includes a set of conditional attributes. These include frequency offset metrics, harmonic distortion metrics, noise level metrics, data quality metrics, residual consistency metrics, and computational load metrics, as well as decision attributes. It is a hybrid mode consisting of sample interpolation recovery mode, direct parameter reconstruction mode, and parameter reconstruction plus residual correction.

[0127] Each conditional attribute is first normalized to the 0-1 range according to the protection setting, health window statistics, or rated resources of the device, and then mapped to three categories of fuzzy semantic quantities: low, medium, and high. For positive risk indicators, the low membership function takes the value 1 when it is less than the first threshold and decreases linearly between the first and second thresholds; the high membership function increases linearly between the third and fourth thresholds and takes the value 1 when it is greater than the fourth threshold; the medium membership function uses a triangular or trapezoidal function in the middle range; data quality indicators are mapped in the opposite direction. The above thresholds are determined by the protection setting or the mean and standard deviation of the health window before the fault.

[0128] When reducing attributes, arbitrarily select a subset of candidate conditional attributes. and the Decision class corresponding to the refactoring pattern For the domain Samples in , First, by attribute Calculate one by one The minimum value is then taken to obtain the conditional similarity. Then click on all those that do not belong to samples Pick The lower bound yields the lower approximate membership degree. According to samples Pick The supremum of the upper bound is obtained to approximate the membership degree. ; The largest among all decision categories Not less than the positive threshold The samples were incorporated , and according to Calculate attribute dependencies. The reduction process starts with an empty attribute set, and each time an attribute is added, it makes the attribute dependency... Increase the maximum conditional attribute until... Achieving full attribute set dependency With preset margin Within the range; if a certain conditional attribute is deleted... If the decrease is less than the preset reduction threshold, the conditional attribute is removed as a redundant indicator. The pattern confidence is obtained by weighting the fuzzy rules corresponding to the reduced attributes, and the pattern corresponding to the highest confidence is taken as the pattern decision result.

[0129] 1) When the device is working in the sample interpolation recovery mode, it directly uses the downsampled sample value in the low speed data area for interpolation recovery, and obtains the reference waveform before the fault according to formula (10).

[0130] Specifically, at the reference time before the failure Select at least four low-velocity historical sampling points nearby, and rank each sampling point according to its relationship with... Construct a cubic Lagrange interpolation function or a local spline interpolation function based on the time distance, and obtain... Reference waveform before the fault occurred.

[0131]

[0132] around Four adjacent low-speed samples were selected nearby, and cubic interpolation was used to recover the health reference value of the corresponding channel. This method handles the shortest link and is suitable for scenarios where the parameter estimation link is temporarily unavailable or device resources are limited.

[0133] In equation (10), Reference time before the failure Health reference values ​​at the location, For interpolation operators, For low-speed data area at time Saved historical sample values.

[0134] 2) When the device is operating in the direct parameter reconfiguration mode, it only uses the parameterized steady-state information in the low-speed data area to construct the reference waveform before the fault, and directly reconfigures it according to Equation (11).

[0135] Parameterized steady-state information is a set of parameters used to describe the healthy steady-state waveform before a fault; the reference waveform before the fault. The parameterized health waveform is reconstructed solely from this set of parameters; the reference waveform before the fault. This is the final health reference value used to subtract the original high-frequency sample value after the fault. In direct parameter reconstruction mode, and Same; in the hybrid mode of parameter reconstruction plus residual correction, Depend on The residual correction is obtained by superimposing the residual correction amount.

[0136]

[0137] Take the fundamental amplitude Third harmonic amplitude Fifth harmonic amplitude Fundamental phase Third harmonic phase Fifth harmonic phase DC bias Substituting the above parameters and τk into equation (11), we obtain the reference time before the fault. Pre-fault reference waveform This method accesses fewer low-speed historical samples, making it easier to quickly obtain the baseline waveform before the fault under conditions of limited computing power.

[0138] In equation (11), for The reference waveform before the fault at any given time, i.e., the parameterized healthy waveform obtained solely from parameterized steady-state information. and The first The amplitude and phase of the subharmonics For the harmonic orders involved in the reconstruction, This is a real-time frequency estimate. DC bias.

[0139] 3) When the device operates in a hybrid mode of parameter reconstruction and residual correction, such as Figure 6 As shown, a reference waveform is first constructed based on parameterized steady-state information, and then a residual is constructed based on the difference between the low-speed sample and the reference waveform. The residual is then used to correct the reference waveform, and the process is carried out according to Equation (12) and Equation (13) respectively.

[0140] Specifically, first in low-speed historical samples Calculate the reference waveform Then calculate the residuals. Subsequently, at the reference time before the failure Nearby, local least squares fitting, local spline interpolation, or piecewise fitting are performed on the residual sequence to obtain... Residual correction at the location The steps for estimating residual corrections include: First, in Before and after the choice satisfies For effective and First, candidate residuals that are not lower than a preset lower limit are used to form a candidate residual set. Second, the candidate residuals are segmented and marked according to residual amplitude, local variance, adjacent residual slope, and source fingerprint of the sampled data. Third, for normal fluctuation residuals, a time distance and... A local quadratic least squares fitting method is used to jointly determine the weights, or local spline interpolation is used when the residual continuity is high, to obtain... Fourth, reduce the weight of distortion residuals and noise-exceeding residuals or remove them from the candidate set, and freeze the output of unusable residuals by using the previous effective correction amount or zero correction amount.

[0141]

[0142]

[0143] around A set of low-velocity historical samples was extracted from the vicinity, and the residual sequence was calculated. Local spline interpolation was then used to obtain the result. Thus, the corrected health reference values ​​are obtained. If we take the offline high-precision health reference value of 0.7211 pu as a control, the absolute errors of the sample interpolation recovery mode, the direct parameter reconstruction mode, and the hybrid mode of parameter reconstruction plus residual correction are approximately 0.0029 pu, 0.0092 pu, and 0.0008 pu, respectively. It can be seen that the hybrid mode of parameter reconstruction plus residual correction has higher reconstruction accuracy when frequency offset and harmonics coexist.

[0144] When there are insufficient available residual points, the device degenerates into low-order interpolation or direct parameter reconstruction mode.

[0145] In hybrid mode, the system processes the residual sequence. Segmented feature labeling is performed. Specifically, residual segments can be labeled as normal fluctuation residuals, distorted residuals, noise-over-limit residuals, or unusable residuals based on residual amplitude, local variance, adjacent residual slopes, fitting error, and channel status indicators. For normal fluctuation residuals, local splines, local least squares, or piecewise fitting estimation are used. For distorted residuals, noise-exceeding residuals, or unusable residuals, the fitting weight is reduced for distorted residuals, noise-exceeding residuals are removed from the local fitting candidate set, and the previous effective correction is frozen or zero correction is applied for unusable residuals. The residual segmentation results can also be fed back to the mode selection module to determine whether to continue using the hybrid mode or degenerate into the sample interpolation recovery mode or the direct parameter reconstruction mode.

[0146] In equation (12), The residual of the low-speed historical sample value relative to the reference waveform; in equation (13), For reference time before the failure The residual correction amount obtained from the estimation, This is the corrected pre-fault health reference value.

[0147] VI. After completing the reconstruction of the reference waveform before the fault, subtract the original high-frequency sample value at the same moment after the fault in the high-speed data area from the healthy reference value to obtain the transient fault component. The calculation formula is shown in equation (14):

[0148]

[0149] If the high-speed sample value after the fault The transient fault component obtained when using a hybrid mode of parameter reconstruction and residual correction By performing the same calculations sequentially on the voltage and current channels of phases A, B, and C (a total of six channels), a complete three-phase transient fault component sequence can be obtained. The fault components can be further constructed into phase mode transformation quantities, zero-sequence quantities, or line modes, and input into the direction discrimination unit, differential discrimination unit, phase selection and ranging unit, and waveform recording and fault diagnosis unit.

[0150] Specifically, in terms of engineering implementation, such as Figure 8 As shown, Figure 2 The functional modules can be collaboratively deployed on FPGA and DSP or CPU according to processing latency and computational characteristics. High-speed sampling, ring buffer management, and fault initiation logic can be preferentially deployed on the FPGA; frequency estimation, mode selection, residual reconstruction, and fault component post-processing can be deployed on the DSP or CPU, thus forming a heterogeneous pipelined implementation architecture suitable for ultra-high-speed protection devices. This invention adopts a hybrid mode of parameter reconstruction plus residual correction as the main operating mode; when low harmonic content, small frequency offset, and limited device computing resources are detected, it can switch to direct parameter reconstruction mode; when the parameter estimation link is abnormal, the quality of low-speed samples is high, or the system is operating in simplified mode, it can degenerate into sample interpolation recovery mode. Through adaptive mode switching, both computational overhead and extraction accuracy can be balanced under different operating conditions.

[0151] Figure 8Solid arrows indicate the data flow direction, while dashed arrows indicate the deployment or scheduling relationship of the corresponding modules by the FPGA or DSP / CPU. The result output module can output transient fault components or phase modulus, zero sequence quantity and line modulus constructed from transient fault components to the direction discrimination unit, differential discrimination unit, phase selection and ranging unit, waveform recording and fault diagnosis unit.

[0152] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting transient fault components with high accuracy for ultra-high frequency sampling, characterized in that, include: S1. Obtain the original high-frequency sample sequence of multiple channels, save the original high-frequency sample sequence of multiple channels in the preset high-speed data area and low-speed data area, and downsample or resample the original high-frequency sample sequence of multiple channels to obtain the health history information before the fault and save it in the low-speed data area; the health history information before the fault includes the steady-state history information before the fault, parameterized steady-state information, data quality indicators of each channel and channel status flags; S2. Perform online frequency measurement based on the health history information before the fault to obtain the real-time frequency estimate and period estimate corresponding to the fault start time; S3. Based on the period estimate, backtrack any high-frequency sampling time after the fault and superimpose the time alignment compensation to construct the reference time before the fault corresponding to the high-frequency sampling time. S4. Based on the steady-state history information before the fault, adaptively select the reconstruction mode and reconstruct the reference waveform before the fault at the reference time before the fault. S5. Subtract the high-speed original sample value at the high-frequency sampling time after the fault from the reference waveform at the reference time before the fault to obtain the transient fault component.

2. The method for extracting accurate and robust transient fault components for ultra-high frequency sampling according to claim 1, characterized in that, S1 includes: S11. A unified sampling clock is used to trigger the analog-to-digital conversion of each channel, and the multi-channel voltage and current analog quantities of the protected object are sampled synchronously at ultra-high frequency to form the multi-channel original high-frequency sample vector and its timestamp corresponding to the current sampling sequence number. S12. Save the original high-frequency sample vectors of the multi-channel in the circular buffer of the high-speed data area. The write pointer increments by taking the modulo of the length of the circular buffer after each new sample vector arrives. After the circular buffer is full, the oldest original high-frequency sample vector is overwritten. S13. After performing anti-aliasing low-pass filtering on the original high-frequency sample vectors of the multi-channel, extract or resample them according to the preset downsampling ratio to obtain low-speed historical sample values ​​and save them in the low-speed data area. S14. Periodically update the parameterized steady-state information of each channel based on the low-speed historical sample values ​​and save it in the low-speed data area. The parameterized steady-state information includes at least one of frequency, amplitude, phase, main harmonic coefficient and DC bias. S15. Save the channel status flags and data quality indicators generated for each channel in the high-speed data area and the low-speed data area. The channel status flags are used to characterize whether there is a disconnection, saturation, sampling link abnormality, sudden distortion, or noise exceeding the limit in the channel. Generate a sampling data traceability fingerprint for each set of original high-frequency samples and save it in the low-speed data area. The sampling data traceability fingerprint includes the sampling link number, analog-to-digital conversion saturation flag, clock node deviation, synchronization pulse jitter, environmental interference intensity, correlation between adjacent channels, and historical anomaly type.

3. The method for extracting accurate and robust transient fault components for ultra-high frequency sampling according to claim 1, characterized in that, S2 includes: S21. Using the main tracking channel as the object, digital phase-locked loop is used to obtain the initial values ​​of the fundamental phase angle and angular frequency; S22. Within the sliding health data window, recursive least squares estimation is used to correct the parameters of the fundamental wave model to obtain the fundamental wave amplitude, phase correction amount, and DC bias. S23. Update the frequency estimate once every preset update interval, and write the updated frequency estimate, fundamental amplitude, phase correction and DC bias into the parameterized steady-state information of the low-speed data area. S24. Based on the change in phase correction at adjacent update times, the base frequency given by the digital phase-locked loop is corrected to obtain the real-time frequency estimate. S25. At the moment of fault initiation, determine the corresponding cycle estimate based on the real-time frequency estimate.

4. The method for extracting accurate and robust transient fault components for ultra-high frequency sampling according to claim 3, characterized in that, The formula for obtaining the real-time frequency estimate is as follows: ； In the formula, Represents the phase expansion operator; Indicates the first The real-time frequency estimate obtained after correction of the base frequency and phase increment at the next update; Indicates the first The base frequency of the digital phase-locked loop output at the next update; This represents the phase correction amount obtained from the recursive least squares estimation; Indicates the first Next update time.

5. The method for extracting accurate and robust transient fault components for ultra-high frequency sampling according to claim 3, characterized in that, S22 includes: S221. Obtain the low-speed historical sample value in the sliding health data window as the data source for recursive least squares estimation. The low-speed historical sample value is the historical sample value after anti-aliasing filtering, abnormal channel elimination and timestamp correction before the fault starts. S222. Preprocess the low-speed historical samples, including removing the initial DC bias value, assigning weights to the samples according to the data quality index, and removing samples marked as unusable by the channel status flag. S223. Based on the preprocessed low-speed historical samples, using the fundamental wave model containing cosine, sine and DC bias terms as the framework, and aiming to minimize the weighted squared residual, the parameters of the fundamental wave model are recursively estimated by recursive least squares update with forgetting factor, and the recursively obtained parameters are converted into fundamental wave amplitude, phase correction and DC bias.

6. The method for extracting accurate and robust transient fault components for ultra-high frequency sampling according to claim 5, characterized in that, S223 includes: S2231. Set the initial covariance matrix and the forgetting factor; the initial covariance matrix is ​​the product of a scalar and an identity matrix, the scalar is selected based on the initial parameter uncertainty, the health window length and the noise level, and the forgetting factor is selected based on the noise level of the health window before the fault and the required tracking speed; S2232. Construct a regression vector and a parameter vector; the regression vector consists of a cosine term, a sine term, and a constant term corresponding to the current sample value; the parameter vector consists of the parameters to be estimated from the fundamental wave model; the phase angle in the cosine and sine terms is obtained by integrating the initial phase angle and angular frequency given by the digital phase-locked loop. S2233. Obtain each preprocessed historical sample value within the sliding health data window according to the timestamp, obtain the weight of each sample value by normalization of the sample value quality factor, and remove the sample points marked as unusable by the channel status flag. S2234. Construct an objective function with the goal of minimizing the weighted squared residual with forgetting factor, and update the objective function with the reduction of the weighted residual of the objective function as the parameter of the recursive process. S2235. Calculate the current recursive gain based on the regression vector, new input sample and its weight corresponding to the current recursive update index, and in combination with the parameter vector and covariance matrix of the previous time step. S2236. Update the parameter vector at the current time step based on the current recursive gain, the parameter vector at the previous time step, and the estimation error of the new input sample. S2237. Based on the current recursive gain, the covariance matrix of the previous time step, and the current regression vector, update the covariance matrix of the current time step. The updated covariance matrix is ​​used for the next recursive gain calculation and parameter reliability judgment. S2238. Calculate the fundamental amplitude and phase correction based on the cosine and sine coefficients in the parameter vector at the current time, and obtain the DC bias based on the DC term in the parameter vector at the current time.

7. The method for extracting accurate and robust transient fault components for ultra-high frequency sampling according to claim 6, characterized in that, The formula for constructing the objective function with the goal of minimizing the weighted squared residual with forgetting factor is as follows: ； In the formula, Indicates the first One historical sample value; Indicates the first The weight of each historical sample value; Indicates the first The regression vector corresponding to each historical sample value; Indicates the parameter estimate; Represents a parameter vector; Indicates the first Objective function for the weighted sum of squared residuals with forgetting factor at each update; Indicates the first The historical sample value at the th The time decay weight in the next update.

8. The method for extracting accurate and robust transient fault components for ultra-high frequency sampling according to claim 1, characterized in that, S4 includes: S41. Calculate the frequency offset index, harmonic distortion index, noise level index, data quality index, residual consistency index, and computational load index respectively. S42. Based on the calculation results of each index, adaptively select any reconstruction mode; if the frequency offset and harmonic distortion are small and computing resources are scarce, select the direct parameter reconstruction mode; if the parameter estimation link is unavailable but the quality of low-speed historical samples is high, select the sample interpolation recovery mode; if the frequency offset, harmonic distortion or waveform distortion are significant and the residual sequence is continuous, select the hybrid mode of parameter reconstruction plus residual correction. S43. When the sample interpolation recovery mode is selected, several low-speed historical sample points are selected near the reference time before the fault, and the interpolation method is used to recover the reference waveform before the fault at the reference time before the fault. S44. When the parameter direct reconstruction mode is selected, the reference waveform before the fault at the reference time before the fault is reconstructed based on the parameterized steady-state information in the low-speed data area. The parameterized steady-state information includes at least one of frequency, amplitude, phase, harmonic coefficient and DC bias. S45. When the hybrid mode of parameter reconstruction plus residual correction is selected, the reference waveform before the fault is reconstructed according to the parameterized steady-state information, and the residual sequence is constructed according to the difference between the low-speed historical sample and the reference waveform. The residual correction amount is estimated near the reference time before the fault. The reference waveform and the residual correction amount are superimposed to obtain the reference waveform before the fault at the reference time before the fault.

9. The method for extracting accurate and robust transient fault components for ultra-high frequency sampling according to claim 8, characterized in that, The formula for reconstructing the pre-fault reference waveform based on parameterized steady-state information is as follows: ； In the formula, This represents the baseline waveform before the fault at time t; Indicates the first Second harmonics; Indicates the first The amplitude of the second harmonic; Indicates the first The phase of the subharmonic; Indicates the harmonic order involved in the reconstruction; This represents the real-time frequency estimate; Indicates DC bias.

10. The method for extracting accurate and robust transient fault components for ultra-high frequency sampling according to claim 8, characterized in that, The estimation of the residual sequence near the reference time before the fault to obtain the residual correction amount includes: Before and after the reference time before the fault, select residual candidate points that meet the requirements of valid channel status flag and data quality index not lower than the preset lower limit to form a residual candidate set. The residual candidate points are segmented and labeled according to the residual amplitude, local variance, adjacent residual slope, and source fingerprint of the sampled data. For candidate points marked as normal fluctuation residuals, local quadratic least squares fitting with weights determined by time distance and data quality indicators is used, or local spline interpolation is used when the residual continuity meets the preset conditions to obtain the residual correction amount at the reference time before the fault. Candidate points marked as distorted residuals and noise-exceeding residuals are reduced in weight or removed from the residual candidate set. For unusable residuals, the previous effective correction amount or zero correction amount is used as the current residual correction amount.

Images