A method for identifying waveform anomalies in high-voltage DC capacitive load systems based on template matching

Abstract

The present invention relates to the field of image processing technology, and in particular to a method for identifying waveform anomalies in a high-voltage DC capacitive load system based on template matching. The method comprises: obtaining a multi-scale waveform of a voltage signal, and preliminarily matching the waveform to be identified with the template waveform at a coarse scale to determine the overall trend. Then, at a medium scale, waveform segmentation is performed, and abnormal segments of the waveform to be identified are screened out, and the corresponding sampling points of the abnormal segments and the corresponding sampling points in the template waveform are mapped and marked to the original scale. Finally, dynamic adaptive time regularization matching is performed based on the waveform to be identified at the original scale and the marked sampling points in the template waveform, and whether the waveform is abnormal is determined based on the matching results. This method avoids local distortion interference during the waveform matching process, and can determine whether the original waveform to be identified is abnormal based on accurate matching results. This method improves the accuracy of waveform anomaly recognition.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method for identifying waveform anomalies in a high-voltage direct current capacitive load system based on template matching. Background Art

[0002] High-voltage DC capacitive load systems are commonly used in critical scenarios such as power transmission, industrial power supply, and renewable energy access. Their operating status is directly related to the safety and stability of the power supply system. The voltage and current waveform signals generated by high-voltage DC capacitive load systems often exhibit complex dynamic characteristics characterized by non-stationary, multi-scale, and multi-frequency components, influenced by multiple factors such as the load's charge and discharge characteristics, equipment aging, electromagnetic interference, and switching operations. These waveforms, as visual representations of time-series data, not only provide a comprehensive picture of system operation but also reveal potential fault trends. Accurately identifying anomalies in these waveforms is key to achieving system health assessment, fault warning, and intelligent operation and maintenance.

[0003] In recent years, with the development of image processing and pattern recognition technology, some studies have proposed converting waveform signals into structured images and identifying anomalies through template matching. This type of method usually uses a preset standard waveform as a template waveform, and performs image-level matching and comparison between the waveform to be identified collected in real time and the template waveform to analyze whether there are waveform anomalies.

[0004] Dynamic time warping (DTC) is one of the most commonly used alignment algorithms in existing template matching technologies. However, in high-voltage DC capacitive load systems, waveform signals are often affected by local disturbances such as spikes and glitches, which can easily cause local distortion. Traditional DTC algorithms are highly sensitive to these local distortions, which can lead to anchoring deviations in the matching path, causing distortion or lag in the overall minimum-cost path. This can lead to mismatches and misidentifications, reducing the accuracy of waveform anomaly detection.

[0005] Therefore, in order to solve the problem of local disturbance in the waveform of high-voltage DC capacitive load system, it is urgent to design a template matching method with global structure perception ability and accurate template matching, which can effectively suppress the interference of local distortion on the matching process, thereby improving the accuracy and stability of waveform anomaly recognition. Summary of the Invention

[0006] To address the problem that traditional template matching methods are sensitive to local distortion in waveform anomaly identification of high-voltage DC capacitive load systems, which easily leads to mismatching and affects the accuracy of waveform anomaly identification, the present invention proposes a method for waveform anomaly identification of high-voltage DC capacitive load systems based on template matching. The method includes:

[0007] Construct the original waveform to be identified and the original template waveform of the voltage signal of the high-voltage DC capacitive load system, and obtain the medium-scale and coarse-scale waveform to be identified and the template waveform through multi-scale downsampling;

[0008] A preliminary match is performed between the coarse-scale waveform to be identified and the template waveform to determine whether the overall waveform trends are consistent; if not, the original waveform to be identified is judged to be abnormal; if consistent, the sampling points corresponding to the coarse-scale waveform to be identified and the template waveform are mapped and marked to the medium-scale waveform to be identified and the template waveform;

[0009] In the mesoscale waveform to be identified and the template waveform, each two adjacent marked sampling points form a segment, and the abnormal segments of the mesoscale waveform to be identified are screened, and the sampling points corresponding to the abnormal segments are mapped and marked to the original waveform to be identified. At the same time, the sampling points corresponding to the abnormal segments in the mesoscale template waveform are mapped and marked to the original template waveform;

[0010] Dynamic time warping matching is performed based on the marked sampling points in the original waveform to be identified and the marked sampling points in the original template waveform, and whether the error of the matching result exceeds the normal error range is determined to determine whether the original waveform to be identified has an abnormality.

[0011] This technical solution first constructs the original waveform and introduces a multi-scale downsampling strategy to extract waveform features at different frequencies and different structural scales from the perspective of signal processing. It can effectively separate high-frequency noise and low-frequency trends, making subsequent analysis more hierarchical. Then, by performing preliminary matching on the coarse-scale waveform, it can quickly determine whether the overall trend of the waveform is consistent, so as to determine whether there is an obvious trend anomaly in the waveform to be identified. At the same time, the structural characteristics of the overall trend are propagated to the mesoscale waveform to achieve structural anchor point transfer and maintain the continuity of the matching logic. Subsequently, a segmented anomaly recognition mechanism is introduced in the mesoscale waveform, and the abnormal segmented features are mapped to the original scale waveform, so that dynamic time regularization is only performed in the abnormal area. Through this waveform matching path convergence mechanism guided by a multi-level structure, the probability of the matching path being interfered with by local distortion is reduced, thereby improving the accuracy of subsequent waveform anomaly recognition.

[0012] Furthermore, the dynamic time warping matching includes: constructing a direct distance matrix based on the direct distance between the marked sampling points in the original waveform to be identified and the marked sampling points in the original template waveform; based on the direct distance matrix, using the dynamic programming method to calculate the cumulative distance matrix, starting from the lower right corner of the cumulative distance matrix and tracing back the path to the upper left corner to obtain the minimum cumulative cost path to complete the dynamic time warping matching; each element of the cumulative distance matrix corresponds to a marked sampling point combination, and the value of the element is the cumulative distance of the marked sampling point combination; the two marked sampling points contained in the marked sampling point combination come from the original waveform to be identified and the original template waveform respectively; the minimum cumulative cost path consists of multiple path points, and each path point corresponds to an element.

[0013] Furthermore, the method for screening abnormal segments of the mesoscale waveform to be identified is: obtaining the corresponding segment of each segment of the mesoscale waveform to be identified in the mesoscale template waveform; determining the abnormality degree of each segment based on the ratio of the cumulative energy of each segment and its corresponding segment and the phase consistency; obtaining the abnormality degree of all segments of the mesoscale waveform to be identified, and using the quartile method to select the segments with abnormality degrees greater than the third quartile as abnormal segments in the mesoscale waveform to be identified.

[0014] This technical solution measures the energy deviation of each segment by comparing its cumulative energy to that of its corresponding segment, indicating whether the segment is experiencing abnormal signal attenuation or amplification. Introducing phase consistency as a supplementary metric further analyzes the phase synchronization between segments, ensuring the identification of potential anomalies when waveform phase changes occur. This approach comprehensively considers both energy deviation and phase synchronization, enabling more comprehensive identification of abnormal segments.

[0015] Furthermore, the method for determining whether the original waveform to be identified has an abnormality is:

[0016] If the error of the matching result exceeds the normal error range, it is determined that the original waveform to be identified has an abnormality; if the error of the matching result does not exceed the normal error range, it is determined that the original waveform to be identified has no abnormality.

[0017] Furthermore, in the process of calculating the cumulative distance matrix using the dynamic programming method, the distance weight coefficient is introduced to reconstruct the recursive formula of the cumulative distance as follows:

[0018]

[0019] In the formula, and are the first The first marked sampling point and the first The direct distance and cumulative distance of the marked sampling points, is the first The first marked sampling point and the first The distance weight coefficient of the marked sampling points, To take the minimum value.

[0020] This technical solution effectively adjusts the matching process of the dynamic time warping algorithm by introducing a distance weight coefficient into the recursive formula for cumulative distance, making path matching more flexible and accurate. Specifically, the cumulative distance coefficient reflects the importance of the waveform carried by each sampling point during the matching process. During the matching process, sampling points with large fluctuations or anomalies are given higher weights, which helps to suppress the impact of noise or unimportant disturbances on the matching path, ensuring that the algorithm still prioritizes the matching path that best conforms to the global trend even in the presence of local distortions.

[0021] Furthermore, the abnormality level of each segment is determined based on the following formula:

[0022] ;

[0023] In this formula, is the first mesoscale waveform to be identified The abnormality level of each segment, is the first mesoscale waveform to be identified The first segment and the mesoscale template waveform The ratio of the cumulative energy of the segments, is the first mesoscale waveform to be identified The first segment and the mesoscale template waveform The phase consistency of each segment, To prevent 0 parameters.

[0024] Furthermore, the distance weight coefficient is determined based on the following formula:

[0025] ;

[0026] In the formula, is the first The first marked sampling point and the first The distance weight coefficient of the marked sampling points, is the original waveform to be identified No. The normalized value of the local signal variance of the marked sampling points, is the original template waveform No. The normalized value of the local signal variance of the marked sampling points.

[0027] This technical solution realizes a dynamic quantitative evaluation of the importance of the corresponding sampling points in the waveform to be identified and the template waveform by introducing a comprehensive weight based on the normalized value of the local signal variance. The local waveform of the waveform is quantified by the local variance. By combining the product of the normalized values ​​of the local variances of the two sampling points and their complementary terms, a confidence measure in a probabilistic sense is formed. The weight avoids the risk of misjudgment caused by a single variance indicator and increases the focus on key feature areas.

[0028] Furthermore, the ratio of the cumulative energy of each segment and its corresponding segment is determined based on the following method: summing the squares of the amplitudes of all sampling points contained in each segment of the mesoscale waveform to be identified as the cumulative energy of the segment; summing the squares of the amplitudes of all sampling points contained in the corresponding segment of the segment in the mesoscale template waveform as the reference cumulative energy of the segment; and taking the ratio of the cumulative energy of the segment to the reference cumulative energy of the segment as the ratio of the cumulative energy of the segment and its corresponding segment, which is used to measure the degree of energy deviation of the waveform to be identified on the segment.

[0029] Furthermore, the phase consistency is determined based on the following method: performing Hilbert transform on the mesoscale waveform to be identified and the template waveform respectively to obtain the instantaneous phase of each marked sampling point of the mesoscale waveform to be identified and the template waveform; determining the phase consistency of each segment of the mesoscale waveform to be identified and its corresponding segment by the difference between the instantaneous phase of the marked sampling point of each segment of the mesoscale waveform to be identified and the instantaneous phase of the marked sampling point of the corresponding segment in the mesoscale template waveform.

[0030] This technical solution, by introducing phase congruence analysis, provides a more accurate means of identifying waveform anomalies in high-voltage DC capacitive load systems. Phase congruence effectively reflects the synchronization of waveforms along the time axis, which is significantly affected when the waveform is disturbed or distorted. This method can identify abnormal waveforms with significant deviations from time synchronization, thereby improving the accuracy of waveform matching and anomaly identification.

[0031] Furthermore, the local signal variance is determined according to the following method: with each marked sampling point as the center, the variance of the amplitudes of all sampling points between the marked sampling point and the marked sampling point closest to its left is recorded as the first signal variance, the variance of the amplitudes of all sampling points between the marked sampling point and the marked sampling point closest to its right is recorded as the second signal variance, and the average of the first variance and the second variance is taken as the local signal variance of the marked sampling point.

[0032] This technical solution can effectively capture the changes in local fluctuations by introducing a calculation method for local signal variance, so that in the waveform matching process, higher weights can be given to local outliers and signal change areas, thereby more accurately reflecting the true characteristics of the waveform and providing more precise local feature identification for waveform matching.

[0033] The present invention has the following effects:

[0034] The present invention adopts a multi-scale collaborative mechanism, template matching concept and local anomaly focusing strategy, and optimizes the multi-scale feature extraction and local anomaly detection of waveforms, thereby avoiding the sensitivity of traditional dynamic time warping algorithms to local distortion in the process of waveform anomaly recognition, ensuring the accuracy of the minimum cumulative cost path, and thus improving the accuracy of waveform anomaly recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic flow chart of the method of the present invention;

[0036] Figure 2 is a schematic flow chart of the method of step S3 of the present invention;

[0037] Figure 3 It is a schematic flow chart of the method of step S4 of the present invention. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0039] The present invention provides a method for identifying abnormal waveforms of high-voltage DC capacitive load systems based on template matching, such as Figure 1 As shown in , including:

[0040] S1: Obtain multi-scale waveforms based on a high-voltage DC capacitive load system.

[0041] In high-voltage DC capacitive load systems, the voltage waveform has complex time-varying characteristics. Influenced by factors such as the capacitive load's charge and discharge characteristics, switching behavior, electromagnetic interference, and equipment aging, the waveform often exhibits abnormal variations at multiple scales and frequency bands. To effectively capture these abnormal characteristics and improve recognition accuracy and robustness, this step performs multi-scale sampling on the waveform signal, taking into account both the global trend and local details of the signal.

[0042] In a high-voltage DC capacitive load system, to analyze whether the voltage waveform within a certain period of time is abnormal, it is necessary to obtain the waveform. Specifically, it includes:

[0043] Assume that the time period is T, that is, the time period includes T acquisition moments, each moment corresponds to an amplitude, and the magnitude of the amplitude represents the magnitude of the voltage. First, obtain the waveform in the time period as the original waveform to be identified, recorded as In this invention, T is set to 1 minute (experienced value, can be adjusted by yourself). At the same time, the waveform of the high-voltage DC capacitive load system in normal operation within a time period of length T is obtained in advance as the original template waveform, recorded as The original waveform to be identified and the original template waveform contain all the detailed information of the waveform, such as high-frequency noise, slight distortion and instantaneous jump.

[0044] The original waveform to be identified and the original template waveform are respectively subjected to multi-scale downsampling, including: downsampling the original waveform to be identified and the original template waveform at a ratio of 2:1 (empirical value), while retaining some details of the waveform, reducing the complexity of the original waveform, and used to detect abnormalities in the intermediate frequency components, such as periodic distortion, regular fluctuations, etc., to obtain the mid-scale waveform to be identified and the mid-scale template waveform, which are respectively recorded as and At the same time, the original waveform to be identified and the original template waveform are downsampled by 5:1 (empirical value), discarding the high-frequency details and noise in the original waveform to be identified and the original template waveform, and retaining the overall contour and trend characteristics of the waveform, such as rising and falling trends, peak positions, periodic trends, etc., to obtain the coarse-scale waveform to be identified and the coarse-scale template waveform, which are respectively recorded as and .

[0045] In signal processing, each point on a waveform is typically referred to as a sampling point. Each sampling point in the original template waveform is assigned a corresponding serial number based on the order in which it was acquired. During downsampling, although some sampling points are discarded, each remaining sampling point originates from the original waveform, and therefore its corresponding serial number in the original waveform can be recorded. This allows a mapping between sampling points in waveforms of different scales to be established.

[0046] S2: Preliminarily judge the overall trend of the waveform based on the coarse-scale waveform.

[0047] The coarse-scale waveform to be identified is transformed using the traditional dynamic time warping algorithm and coarse-scale template waveform Match them to get the minimum cumulative cost path between the two. Each path point on the path corresponds to a sampling point combination. The two sampling points contained in the sampling point combination come from and The direct distance between the two sampling points (the direct distance is the absolute value of the difference between the amplitudes of the two sampling points) is used as the matching error for that path point. The matching error of each path point is normalized by the maximum and minimum values ​​to obtain the normalized matching error for each path point. The coefficient of variation of the normalized matching error of all path points (the ratio of the standard deviation of the normalized matching error to the mean) is calculated to reflect the relative fluctuation of the normalized matching error.

[0048] Calculate the comprehensive matching error of the matching path , is the mean of the normalized matching errors of all path points on the path, reflecting the average error level. The coefficient of variation of the normalized matching error for all path points reflects the fluctuation level of the normalized matching error. A larger mean indicates a larger matching error, and a smaller coefficient of variation indicates a higher confidence level for a larger matching error, and vice versa.

[0049] Therefore, the matching error threshold is set to 0.8 (empirical value). , indicating that the overall trend of the original waveform to be identified is consistent with that of the original template waveform and the fluctuation is stable. Further analysis and positioning is needed to determine whether there is a local anomaly. If (Empirical value), indicating that the overall trend of the original waveform to be identified deviates greatly from that of the template waveform, and the waveform to be identified has obvious trend anomalies.

[0050] S3: Map the coarse-scale waveform to the mesoscale waveform and segment it, evaluate the abnormality of each waveform segment, and locate the abnormal waveform segment of the mesoscale waveform.

[0051] In high-voltage DC capacitive load systems, voltage waveform anomalies can arise from a variety of factors, including global trend deviations, local energy mutations, and phase timing disturbances. While the original waveform to be identified contains complete details, its high-frequency noise and complex time-varying characteristics make it difficult to directly use to accurately locate anomalies. While coarse-scale waveforms to be identified can quickly filter out noise and capture global trends, such as overall voltage offsets or charge-discharge cycle anomalies, they are overly smoothed and fail to capture the waveform's mid-frequency details.

[0052] Therefore, this step maps the coarse-scale waveform to be identified to the mesoscale waveform to be identified, marking the sampling points corresponding to the coarse-scale waveform on the mesoscale waveform. These sampling points usually correspond to key nodes of the waveform, such as waveform peaks, zero crossings, or turning points in the charge and discharge phases. By segmenting the mesoscale waveform using these key nodes as boundaries, the continuous waveform can be broken down into independent segments aligned with the physical processes of the system. This operation not only retains the guidance of the coarse-scale waveform on the global trend, but also uses the medium resolution of the mesoscale to capture the mid-frequency details, making each segment an independent segment that is both a global positioning anchor and a local feature carrier. This avoids the global analysis from masking local anomalies, and also breaks through the noise interference of the original waveform, ultimately achieving cross-scale precise analysis from coarse-scale trend screening to mesoscale local positioning.

[0053] Specifically, if Figure 2 Shown, including:

[0054] S31: Map and mark the sampling points contained in the coarse-scale waveform to the medium-scale waveform.

[0055] The sampling points contained in the coarse-scale waveform to be identified and the coarse-scale template waveform are respectively mapped to the mesoscale waveform to be identified and the mesoscale template waveform, and marked to obtain all the marked sampling points in the mesoscale waveform to be identified and all the marked sampling points in the mesoscale template waveform.

[0056] S32: Segment the mesoscale waveform.

[0057] In the mesoscale waveform to be identified, each two adjacent marked sampling points form a segment, completing the segmentation operation of the mesoscale waveform to be identified. The mesoscale template waveform is also segmented, and the segmentation operation method is consistent with the segmentation operation method for the mesoscale waveform to be identified.

[0058] Since the time length of the coarse-scale waveform to be identified and the coarse-scale template waveform are the same, and the sampling points of the two are one-to-one corresponding based on the acquisition time, the coarse-scale waveform to be identified and the coarse-scale template waveform are mapped to the mesoscale waveform and segmented, and the corresponding segment of each segment in the mesoscale waveform to be identified in the mesoscale template waveform is obtained.

[0059] S33: Calculate the ratio of the cumulative energy of each segment and its corresponding segment.

[0060] In high-voltage DC capacitive load systems, the energy distribution of the voltage waveform is closely related to the system's operating state. Under normal operating conditions, the energy characteristics of each stage (such as the energy accumulation rate during charging and discharging, and the energy mutation amplitude during switching) are stable and regular. However, when the system experiences anomalies (such as load leakage, line impedance changes, or equipment aging), the energy in localized intervals may deviate significantly, and these anomalies are often hidden in the waveform's intermediate frequency components.

[0061] Therefore, by calculating the ratio of cumulative energies, the abstract waveform differences are converted into quantifiable energy deviation indicators. If the cumulative energy of a segment of the mesoscale waveform to be identified is significantly higher than the corresponding segment of the mesoscale template waveform, it indicates that the segment is more likely to have energy excess due to decreased capacitive load energy storage efficiency, abnormally increased power supply voltage, or high-frequency harmonic injection. If the cumulative energy of a segment of the mesoscale waveform to be identified is significantly lower than the corresponding segment of the mesoscale template waveform, it indicates that the segment is more likely to have energy loss due to abnormal load energy consumption, poor line contact, or obstructed charging and discharging processes.

[0062] In one embodiment, the ratio of the cumulative energy of each segment and its corresponding segment is determined based on the following formula:

[0063]

[0064] In this formula, is the mesoscale waveform to be identified No. Segmented and mesoscale template waveforms No. The cumulative energy ratio of the segments, for or No. The sequence number of the sampling points contained in each segment, for or No. The total number of sampling points contained in each segment, for No. The first segment The amplitude of the sampling points, for No. The first segment The energy of the sampling points, for No. The accumulated energy of each segment. for No. The first segment The amplitude of the sampling points, for No. The first segment The energy of the sampling points, for No. The accumulated energy of each segment.

[0065] like , represents the first The energy of the first segment is higher than that of the template waveform at the middle scale. There may be energy anomalies or signal enhancement in this segment. :Describe the first part of the mesoscale waveform to be identified The energy of the first segment is lower than that of the template waveform at the middle scale. If the signal is weakened or the information is missing. :Describe the first part of the mesoscale waveform to be identified The first segment and the mesoscale template waveform If the waveform segments are similar, they may be normal or well-matched signals.

[0066] S34: Determine the phase consistency of each segment and its corresponding segment.

[0067] In high-voltage DC capacitive load systems, the phase characteristics of the voltage waveform are a core indicator of the synchronization and stability of the system's dynamic processes. During normal operation, the phase changes of the voltage waveform are strictly synchronized with the charge and discharge cycles of the capacitive load and the operation timing of the switching devices, exhibiting a fixed phase difference or periodic pattern. However, when the system experiences anomalies (such as switching delays, electromagnetic interference coupling, or device parameter drift), the phase relationship of the waveform may become disordered, causing the timing structure between each segment to deviate from the normal pattern. Determining phase consistency is necessary because it can reveal dynamic process anomalies hidden in the waveform from the perspective of signal synchronization, compensating for the limitations of relying solely on energy analysis.

[0068] In one embodiment, phase consistency is determined as follows:

[0069] respectively and Performing Hilbert transform, we can obtain the instantaneous phase of each marked sampling point, that is, the phase information corresponding to the waveform signal at each sampling point, which is an important indicator to measure the phase synchronization of two signals.

[0070] Then, by calculating Each waveform segment and The phase consistency of the two waveform segments is quantified by the phase locking value of the corresponding waveform segment in:

[0071]

[0072] In this formula, for No. Segments and No. The phase consistency of each segment, that is, the phase lock value, is a statistic used to measure signal synchronization; when the phase lock value is 1, the phase difference between the two signals is constant, that is, completely synchronized; when the lock value is 0, the phase is not synchronized. for No. The first segment The instantaneous phase of the sampling points, for No. The first segment The instantaneous phase of the sampling points, represents the imaginary unit, represents a complex number of unit length, the phase of which is , is an imaginary unit, which converts the angle difference into a complex number on the unit circle, and It is through Euler's formula To determine, for or No. The total number of sampling points contained in each segment, To find the modulus length symbol, is a natural constant.

[0073] In this formula, represents a complex number of unit length, whose phase is . It will The instantaneous phase and sampling point of each segment The instantaneous phase differences of the sampling points of the corresponding segments in are added up in complex form and the average modulus is calculated to measure the phase synchronization degree of the two waveform segments.

[0074] In short, when the phase consistency of the two waveform signals is greater, the synchronization is better (the phase difference is a constant), the modulus of the sum of the complex numbers is larger and closer to 1; when the phase is randomly distributed, the modulus of the sum of the complex numbers approaches 0 and closer to 0.

[0075] S35: Determine the degree of abnormality by combining the ratio of phase consistency and cumulative energy.

[0076] Single energy or phase analysis often fails to fully characterize the complex characteristics of waveform anomalies. Energy deviations may reflect sudden amplitude changes or power imbalances, while phase disturbances indicate timing desynchronization or dynamic process distortion. In actual anomaly scenarios, both often occur together (for example, load aging can lead to increased energy loss and may also cause charge and discharge phase shifts). Therefore, combining phase consistency and the ratio of accumulated energy to determine the degree of anomaly avoids the one-sidedness of single feature analysis.

[0077] In one embodiment, the abnormality level is determined based on the following formula:

[0078]

[0079] In this formula, is the first mesoscale waveform to be identified The abnormality level of each segment, is the first mesoscale waveform to be identified The first segment and the mesoscale template waveform The ratio of the cumulative energy of the segments, is the first mesoscale waveform to be identified The first segment and the mesoscale template waveform The phase consistency of each segment, To prevent 0 parameters. The value is .

[0080] In this formula, The first part measures the mesoscale waveform to be identified. The first segment and the mesoscale template waveform The energy deviation degree of each segment is When it is significantly greater than or less than 1, it means that the greater the degree of energy deviation, the more likely it is that it is an energy mutation feature of an abnormal waveform. The purpose of the squaring operation is to strengthen the degree of energy deviation. Whether it is too high or too low, it is considered abnormal (symmetry processing). Part of it is used as a normalization factor to avoid the abnormality being infinitely magnified due to excessive energy deviation. This is to prevent the denominator from being zero or too small, which may cause numerical instability, and to maintain abnormal numerical stability. The first one represents the mesoscale waveform to be identified The first segment and the mesoscale template waveform If the overall phase trend of the segments is different, Close to 1, the phase trends are almost the same, The closer it is to 0, the more likely the segment is a normal waveform and the smaller the abnormality. The closer to 0, if The closer it is to 0, the more different the phase trend is, the more likely the segment is an abnormal waveform, and the greater the degree of abnormality, The closer to 1.

[0081] S36: Filter abnormal segments of the mesoscale waveform to be identified according to the degree of abnormality.

[0082] Get The abnormality level of all segments , sort them in ascending order, and then use the quartile method to classify the waveform segments with an abnormality greater than the third quartile (greater than 75%) as abnormal waveform segments.

[0083] S4: Map the abnormal waveform of the mesoscale waveform to the original scale waveform in segments, perform adaptive dynamic time warping matching, and determine the optimal matching path.

[0084] Although the abnormal segmentation of the mesoscale waveform to be identified has been located in the mid-frequency region where the energy and phase deviate significantly, this region may still contain detailed features such as high-frequency noise and transient jumps in the original waveform, which directly affects the accurate determination of the nature of the anomaly, such as distinguishing between progressive distortion caused by equipment aging and occasional spikes caused by transient interference.

[0085] Therefore, this step maps the abnormal segments of the mesoscale waveform to be identified to the original waveform to be identified and performs adaptive dynamic time warping matching. It can perform a refined analysis of the high-frequency details of the abnormal area while retaining the coarse-scale global constraints and mesoscale positioning accuracy.

[0086] According to the mesoscale waveform to be identified Contains sampling points and original waveform to be identified The mapping relationship of the sampling points included in the middle scale waveform to be identified All sampling points contained in each abnormal segment are mapped and marked on the original waveform to be identified In the example, we get the original waveform to be identified All marked sampling points.

[0087] Similarly, the mesoscale waveform to be identified Each abnormal segment in the mesoscale template waveform All sampling points contained in the corresponding segment in the original template waveform are mapped and marked In the example above, we get the original template waveform All the marked sampling points in .

[0088] Specifically, the process of adaptive dynamic time warping matching is as follows Figure 3 Shown, including:

[0089] S41: Calculate the direct distance matrix between all the marked sampling points of the original waveform to be identified and all the marked sampling points of the original template waveform.

[0090] Get it first Each marked sampling point and The direct distance of each marked sampling point (the difference in the amplitude of the sampling point, and then the absolute value), let The number of marked sampling points is , marking the sampling point number as , The value range is 1 to ; The number of marked sampling points is , marking the sampling point number as , The value range is 1 to ,according to Each marked sampling point to The direct distance between each marked sampling point constitutes a Direct distance matrix of .

[0091] In the direct distance matrix in Elements at The value of No. Marked sampling points and No. The direct distance between the two marked sampling points. The smaller the direct distance, the smaller the matching cost of the two marked sampling points, and vice versa.

[0092] S42: Based on the direct distance matrix, the dynamic programming method is used to calculate the cumulative distance matrix, and in this process, the distance weight coefficient is introduced to reconstruct the recursive formula of the cumulative distance.

[0093] The specific steps include:

[0094] S421: Initialize boundary conditions.

[0095] when hour:

[0096] Cumulative Matrix Origin (upper left corner), and , ;

[0097] Cumulative Matrix First row (horizontally advancing), and , ,in, ;

[0098] Cumulative Matrix The first column (descending vertically), and , , ;

[0099] S422: Dynamically reconstruct the recursive formula of the cumulative distance using the weight coefficient to calculate the cumulative distance matrix.

[0100] when During the recursive process of cumulative distance calculation, traditional recursive formulas typically assign fixed weights to the transfer paths of each time step, which can easily make the algorithm difficult to adapt to dynamic changes in the local characteristics of the waveform, such as sudden amplitude changes, phase jumps, or noise interference. Therefore, by introducing dynamic weight coefficients, the calculation of cumulative distance can dynamically respond to differences in local waveform characteristics, thereby improving the adaptability and accuracy of the matching algorithm for complex signals.

[0101] In one embodiment, a distance weight coefficient is introduced to reconstruct the recursive formula of the cumulative distance as follows:

[0102]

[0103] In this formula, is the first The first marked sampling point and the first The direct distance to the marked sampling points, is the first The first marked sampling point and the first The cumulative distance between the marked sampling points. To get the minimum value, is the first The first marked sampling point and the first The distance weight coefficient of each marked sampling point.

[0104] In this formula, The first step for dynamically and adaptively adjusting the original waveform to be identified The first marked sampling point and the first The direct distance to the marked sampling point The weight distribution of the historical cumulative distance is usually in the range of . The larger the value, the more it relies on direct distance when calculating the cumulative distance to reflect the matching cost between sampling points. , Leading the calculation of cumulative distance, it enhances the accuracy of capturing abnormal features such as amplitude mutation and phase jump; The smaller it is, the more it relies on the historical cumulative distance when calculating the cumulative distance to reflect the matching cost between sampling points. The smooth trend based on the historical cumulative distance can suppress noise interference and maintain the continuity of time series matching.

[0105] In summary, this dynamic adjustment mechanism enables the algorithm to flexibly switch between precise alignment of local features and smooth transition of global trends, improving its adaptability to high-voltage DC capacitive load waveforms containing noise, time-varying or multi-scale anomalies. It not only avoids the omission or mismatching of abnormal features by traditional methods, but also enhances the robustness of the matching process, ultimately providing a more reliable quantitative basis for high-precision waveform anomaly positioning and fault diagnosis.

[0106] In one embodiment, the dynamic weight coefficient is calculated based on the following formula:

[0107]

[0108] In this formula, is the first The first marked sampling point and the first The distance weight coefficient of the marked sampling points, is the original waveform to be identified No. The normalized value of the local signal variance of the marked sampling points is equal to No. The local signal variance of the marked sampling points is divided by The maximum value of the local signal variance of all marked sampling points. is the original template waveform No. The normalized value of the local signal variance of the marked sampling points is equal to No. The local signal variance of the marked sampling points is divided by The maximum value of the local signal variance of all marked sampling points.

[0109] In this formula, the numerator This method measures the consistency of local fluctuations between the sampling points of the original waveform to be identified and the corresponding sampling points of the original template waveform. High local signal variances at both sampling points, and a large product value, indicate a high likelihood of shared anomalies or strong dynamics, necessitating more focused matching. A larger distance weighting coefficient indicates a greater reliance on the direct distance between the two sampling points when performing matching.

[0110] In this formula, the denominator is equal to the sum of the numerator and the complement of the numerator. It is the complementary term of the numerator term, reflecting the local consistency between the sampling points of the original waveform to be identified and the corresponding sampling points of the original template waveform.

[0111] If the first The first marked sampling point and the first The smaller the local signal variance of the marked sampling point, the more stable the waveform. At this time, the smaller the numerator, the larger the complementary term of the numerator in the denominator, and the denominator is dominated by the complementary term of the numerator. At this time, the smaller the distance weight coefficient, the more dependent on the historical cumulative distance is when matching.

[0112] On the contrary, if the first The first marked sampling point and the first The larger the local signal variance of the marked sampling point, the more violent the waveform fluctuation. At this time, the larger the numerator term is, the smaller the complementary term of the numerator term is in the denominator term, and the denominator is dominated by the numerator term. At this time, the larger the distance weight coefficient is, the more dependent on direct distance is when matching.

[0113] In one embodiment, the local signal variance of each marked sampling point is determined based on the following method: for any marked sampling point of the original waveform to be identified or the original template waveform, with the marked sampling point as the center, all sampling points included between the marked sampling point and the nearest marked sampling point on the left (excluding the nearest marked sampling point on the left), and the variance of the amplitudes of these sampling points are calculated; all sampling points included between the marked sampling point and the nearest marked sampling point on the right (excluding the nearest marked sampling point on the right), and the variance of the signal amplitudes of these sampling points are calculated; and the average of these two variances is used as the local signal variance of the marked sampling point.

[0114] The local signal variance of a marker sampling point reflects the importance of that sampling point itself, that is, whether it is unusually active or stable. The purpose is to focus the matching on areas of significant change. A larger local signal variance means that the local signal fluctuations at that marker sampling point are more dramatic, which may correspond to the marker sampling point being more likely to be in a key part of the waveform, thus being more important and therefore being given a higher weight. The opposite is true for the same sampling point.

[0115] Finally, the cumulative distance matrix is ​​obtained by calculating each element of the cumulative distance matrix according to the recursive formula of the reconstructed cumulative distance. Each element of the cumulative distance matrix corresponds to a marked sampling point combination, and the value of the element is the cumulative distance of the marked sampling point combination. The two marked sampling points contained in the marked sampling point combination come from the original waveform to be identified and the original template waveform respectively.

[0116] S43: Perform path backtracking based on the cumulative distance matrix to obtain the optimal matching path.

[0117] Specifically, the path is traced back from the lower right corner of the cumulative distance matrix to the upper left corner to obtain the minimum cumulative cost path, which is used as All marked sampling points and The optimal matching path between all marked sampling points is found to complete the adaptive dynamic time warping matching.

[0118] Specifically, the path backtracking process is:

[0119] S431: Starting from the element at the lower right corner of the accumulation matrix, use this element as the current path point;

[0120] S432: Select the smallest element among the elements above, to the left, and in the upper left corner of the current path point as the next path point.

[0121] For example, if the current path point is , in its adjacent elements 、 or Select an element with the smallest value as the next path point.

[0122] S433: Update the current path point, take the next path point of the current path point as the new current path point, and execute step S432; until the path backtracks to the element in the upper left corner of the accumulation matrix, the path backtracking is completed.

[0123] All path points in the backtracking process constitute a minimum cumulative cost path, which reflects the optimal alignment of the marked sampling points of the original waveform to be identified and the marked sampling points of the original template waveform on the time axis. Therefore, it is used as the optimal matching path of the marked sampling points of the original waveform to be identified and the marked sampling points of the original template waveform. The optimal matching path consists of multiple path points, each path point corresponds to an element, and each element is composed of two marked sampling points.

[0124] S5: Calculate the matching error based on the optimal matching path to perform waveform anomaly recognition.

[0125] In the dynamic time warping algorithm, the optimal matching path is obtained by accumulating a distance matrix, reflecting the minimum matching cost between two sequences. In the present invention, the optimal matching path represents the optimal alignment between the labeled sampling points of the original waveform to be identified and the labeled sampling points of the original template waveform. Each path point on the optimal matching path corresponds to a labeled sampling point combination, and each labeled sampling point combination includes two labeled sampling points, one from the original waveform to be identified and the other from the original template waveform. The direct distance between these two sampling points is used as the matching error corresponding to the path point. The comprehensive matching error of the optimal matching path is calculated with reference to the operation in step S2.

[0126] Specifically:

[0127] The matching error of each path point on the optimal matching path is normalized to the maximum and minimum values ​​to obtain the normalized matching error of each path point. Then, the comprehensive matching error of the optimal matching path is calculated. , is the mean of the normalized matching errors of all path points of the optimal matching path, The coefficient of variation of the normalized matching error for all path points in the optimal matching path. A larger mean indicates a larger matching error, and a smaller coefficient of variation indicates a higher confidence level for a larger matching error, and vice versa.

[0128] Similarly, set the matching error threshold to 0.5 (experience value), if , indicating that the matching between the marked sampling points in the original waveform to be identified and the marked sampling points in the original template waveform is accurate, and it is determined that there is no abnormality in the original waveform to be identified; if (empirical value), indicating that the matching error between the marked sampling points in the original waveform to be identified and the marked sampling points in the original template waveform is too large, and it is determined that the original waveform to be identified is abnormal.

[0129] While various embodiments of the present invention have been shown and described herein, it will be obvious to those skilled in the art that such embodiments are provided by way of example only.

Claims

1. A method for identifying waveform anomalies in a high-voltage DC capacitive load system based on template matching, characterized in that: include: Construct the original waveform to be identified and the original template waveform of the voltage signal of the high-voltage DC capacitive load system, and obtain the medium-scale and coarse-scale waveform to be identified and the template waveform through multi-scale downsampling; Perform a preliminary match between the coarse-scale waveform to be identified and the template waveform to determine whether the overall waveform trends are consistent; if not, it is determined that the original waveform to be identified is abnormal; If they are consistent, the sampling points corresponding to the coarse-scale waveform to be identified and the template waveform are mapped and marked to the medium-scale waveform to be identified and the template waveform; In the mesoscale waveform to be identified and the template waveform, each two adjacent marked sampling points are a segment, and the abnormal segments of the mesoscale waveform to be identified are screened, specifically: Obtaining the corresponding segment of each segment of the mesoscale waveform to be identified in the mesoscale template waveform; determining the abnormality degree of each segment based on the ratio of the cumulative energy of each segment and its corresponding segment and the phase consistency; Obtaining the abnormality levels of all segments of the mesoscale waveform to be identified, and using the quartile method to select segments with abnormality levels greater than the third quartile as abnormal segments in the mesoscale waveform to be identified; The sampling points corresponding to the abnormal segments are mapped and marked to the original waveform to be identified. At the same time, the sampling points corresponding to the abnormal segments in the mesoscale template waveform are mapped and marked to the original template waveform. Dynamic time warping matching is performed based on the marked sampling points in the original waveform to be identified and the marked sampling points in the original template waveform, and whether the error of the matching result exceeds the normal error range is determined to determine whether the original waveform to be identified has an abnormality.

2. The method for identifying waveform anomalies of a high-voltage direct current capacitive load system based on template matching according to claim 1, characterized in that: The dynamic time warping matching includes: A direct distance matrix is ​​constructed based on the direct distances between the marked sampling points in the original waveform to be identified and the marked sampling points in the original template waveform. Based on the direct distance matrix, a cumulative distance matrix is ​​calculated using a dynamic programming method. The path is traced back from the lower right corner of the cumulative distance matrix to the upper left corner to obtain the minimum cumulative cost path to complete dynamic time warping matching. Each element of the cumulative distance matrix corresponds to a marked sampling point combination, and the value of the element is the cumulative distance of the marked sampling point combination; the two marked sampling points contained in the marked sampling point combination come from the original waveform to be identified and the original template waveform respectively; the minimum cumulative cost path is composed of multiple path points, and each path point corresponds to an element.

3. The method for identifying waveform anomalies in a high-voltage direct current capacitive load system based on template matching according to claim 2, wherein: The method for determining whether the original waveform to be identified has an abnormality is: If the error of the matching result exceeds the normal error range, it is determined that the original waveform to be identified is abnormal; If the error of the matching result does not exceed the normal error range, it is determined that there is no abnormality in the original waveform to be identified.

4. The method for identifying waveform anomalies in a high-voltage direct current capacitive load system based on template matching according to claim 2, wherein: In the process of calculating the cumulative distance matrix using the dynamic programming method, the distance weight coefficient is introduced to reconstruct the recursive formula of the cumulative distance as follows: ； In the formula, and are the first The first marked sampling point and the first The direct distance and cumulative distance of the marked sampling points, is the first The first marked sampling point and the first The distance weight coefficient of the marked sampling points, To take the minimum value.

5. The method for identifying waveform anomalies in a high-voltage direct current capacitive load system based on template matching according to claim 1, wherein: The abnormality level of each segment is determined based on the following formula: ； In this formula, is the first mesoscale waveform to be identified The abnormality level of each segment, is the first mesoscale waveform to be identified The first segment and the mesoscale template waveform The ratio of the cumulative energy of the segments, is the first mesoscale waveform to be identified The first segment and the mesoscale template waveform The phase consistency of each segment, To prevent 0 parameters.

6. The method for identifying waveform anomalies in a high-voltage direct current capacitive load system based on template matching according to claim 4, characterized in that: The distance weight coefficient is determined based on the following formula: ； In the formula, is the first The first marked sampling point and the first The distance weight coefficient of the marked sampling points, is the original waveform to be identified No. The normalized value of the local signal variance of the marked sampling points, is the original template waveform No. The normalized value of the local signal variance of the marked sampling points.

7. The method for identifying waveform anomalies in a high-voltage direct current capacitive load system based on template matching according to claim 5, characterized in that: The ratio of the cumulative energy of each segment and its corresponding segment is determined based on the following method: The squares of the amplitudes of all sampling points in each segment of the mesoscale waveform to be identified are summed as the cumulative energy of the segment; The squares of the amplitudes of all sampling points contained in the corresponding segment of the segment in the mesoscale template waveform are summed as the reference cumulative energy of the segment; the ratio of the cumulative energy of the segment to the reference cumulative energy of the segment is used as the ratio of the cumulative energy of the segment and its corresponding segment, which is used to measure the degree of energy deviation of the waveform to be identified in the segment.

8. The method for identifying waveform anomalies in a high-voltage direct current capacitive load system based on template matching according to claim 5, wherein: The phase consistency is determined based on the following method: Performing Hilbert transform on the mesoscale waveform to be identified and the template waveform respectively to obtain the instantaneous phase of each marked sampling point of the mesoscale waveform to be identified and the template waveform; The phase consistency between each segment of the mesoscale waveform to be identified and its corresponding segment is determined by the difference between the instantaneous phase of the marked sampling point of each segment of the mesoscale waveform to be identified and the instantaneous phase of the marked sampling point of the corresponding segment in the mesoscale template waveform.

9. The method for identifying waveform anomalies in a high-voltage direct current capacitive load system based on template matching according to claim 6, wherein: The local signal variance is determined as follows: Taking each marked sampling point as the center, the variance of the amplitudes of all sampling points between the marked sampling point and the nearest marked sampling point on its left is recorded as the first signal variance, and the variance of the amplitudes of all sampling points between the marked sampling point and the nearest marked sampling point on its right is recorded as the second signal variance; the mean of the first signal variance and the second signal variance is taken as the local signal variance of the marked sampling point.

Images