Fault recording data intelligent comparison method based on fast dynamic time warping algorithm

Abstract

This invention relates to an intelligent comparison method for fault waveform recording data based on a fast dynamic time warping algorithm. It addresses the inconsistency in sampling frequencies between the protection device and the fault recorder by performing sampling frequency conversion on the extracted fault waveform recording data. Then, it uses Euclidean distance to align waveforms from the same source to achieve clock synchronization. Finally, the FDTW algorithm is used for intelligent comparison of the fault waveform recording data. This invention can perform consistency processing on waveform recording data from the same source, achieving consistent sampling frequencies and clock synchronization, and realizing similarity matching between waveforms recorded by the protection device and those recorded by the fault recorder in a vacuum semantic environment.

Description

Technical Field

[0001] This invention belongs to the field of substation equipment fault detection technology, and in particular to a method for intelligent comparison of fault recording data based on a fast dynamic time warping algorithm. Background Technology

[0002] The rapid construction of smart substations and their continuous improvement in automation levels have created a favorable environment for the development of fault recording technology. Fault recording data is an important basis for evaluating protection actions and analyzing the nature and causes of related equipment faults. However, in the early stages, there were no relevant fault recording technology specifications in the market, and different manufacturers had different requirements for fault recording technology. This resulted in fault recording files exported from relay protection devices having problems such as lack of channel identification information and lack of branch ratio coefficients, which create a vacuum semantic environment. It is impossible to match fault recording files from the same source and channel with the recording files from protection devices, which increases the difficulty of centralized and regional management of recording substations. Therefore, it is urgent to study waveform matching technology based on vacuum semantic environment to realize intelligent comparison of fault recording data from the same source under abnormal sampling circuit conditions, thereby improving the reliability of power system fault diagnosis.

[0003] Currently, fault waveform matching analysis techniques are mainly divided into direct waveform similarity matching and indirect matching. Direct waveform similarity matching uses time-series data to compare the similarity between two waveforms to be matched, thus determining the matching status. Key methods include the correlation coefficient method, Dynamic Time Warping (DTW), and Hausdorff distance method. The correlation coefficient method is fast and simple, but it requires the waveforms to be matched to have the same length and has poor anti-interference capabilities. DTW does not require a one-to-one correspondence between the data points of the waveforms to be matched, effectively solving the waveform similarity matching problem after deformation. However, when the amount of waveform data to be matched is large, the DTW algorithm suffers from slow computation speed and high computational complexity. The Hausdorff distance method uses the maximum value of the minimum one-way distance between the waveforms to be matched to measure waveform similarity, but this method itself has weak anti-interference capabilities and is easily affected by abnormal sampling points.

[0004] Indirect waveform similarity matching utilizes partial feature quantities to replace the waveform data to be matched, achieving spatial transformation of the time series. The similarity between two waveforms is measured by processing these feature quantities. Major methods include wavelet transform, phase plane method, and S-transform. Among these, the wavelet transform method uses scaling and translation operations to refine related waveforms at multiple scales, effectively highlighting specific characteristics of the problem and enabling localized analysis. The wavelet coefficients for signal noise are relatively small, and noise can be removed through threshold selection. The phase plane method can extract global characteristics of the waveform and is unaffected by phase amplitude. However, it is overly conservative in data processing, has poor fault tolerance, and insufficient anti-interference capability.

[0005] There is a lack of literature on methods for comparing fault waveforms from the same source. These studies do not take into account situations where the sampling files of the protection device lack channel information when the circuit is abnormal, which increases the difficulty of comparing fault waveform files from the same source. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and propose an intelligent comparison method for fault recording data based on a fast dynamic time warping algorithm. This method can perform consistency processing on the same source recording file data, achieve consistent sampling frequency and clock synchronization, and at the same time realize the similarity matching between the waveform of the protection device and the waveform of the fault recorder in a vacuum semantic environment.

[0007] The technical problem solved by this invention is achieved through the following technical solution:

[0008] The intelligent comparison method for fault recording data based on the fast dynamic time warping algorithm includes the following steps:

[0009] Step 1: Extract the fault waveform data that needs to be compared;

[0010] Step 2: Determine whether the sampling frequencies of the fault recording data are consistent. If they are consistent, proceed to Step 4; otherwise, proceed to Step 3.

[0011] Step 3: Convert the sampling frequency of the fault recording data to make the sampling frequency of the fault recording data consistent.

[0012] Step 4: Determine if the time synchronization of the fault recording data is consistent. If it is, proceed to step 6; otherwise, proceed to step 5.

[0013] Step 5: Align the waveforms of the fault recording data for consistency.

[0014] Step 6: Perform waveform similarity matching on the fault recording data and compare and analyze the matching results.

[0015] Furthermore, the specific implementation method of step 3 is as follows: using the Lagrange interpolation resampling method to convert the sampling frequency of the fault recording file data:

[0016] The original fault recording data waveform sampling rate f o Switch to the new sampling frequency f n Calculate the waveform length F of the fault recording data under the new sampling rate:

[0017]

[0018] Where G is the waveform length of the original fault recording data;

[0019] Calculate the specific location c of the interpolation within the sampling interval of the original fault recording data. k :

[0020]

[0021] Where k is the sample value index, and its range is 1≤k≤F; the time exponent c is set (1≤c≤G), and:

[0022] c≤c k ≤c+1

[0023] The window length is determined, where the highest order of the Lagrange interpolation polynomial is 2×v, the size of the window function is 2×v+1, and the time factors for the 2×v+1 original samples are as follows:

[0024] cv,…,c-1,c,c+1,…,c+v

[0025] The resampling weights are calculated for all points within the window length using the Lagrange resampling weight formula:

[0026]

[0027] Where i = -v, -v+1, ..., v are the window positions where the interpolation points are located, and l i Let $l$ be the resampling weight for the interpolation point $i$, and $j$ be the window position of the other interpolation points besides $i$, where $j = -v, -v+1, ..., v$ and $j ≠ i$. Based on the magnitude and resampling weight $l$ of each interpolation point... i Calculate the data amplitude at the interpolation point and obtain the weighted polynomial resampled sequence L(k) of 2×v+1 original sample values:

[0028]

[0029] Where m(ci) is the sampling sequence of the original fault recording data.

[0030] Furthermore, step 5 includes the following steps:

[0031] Step 5.1: Using one of the fault recording waveforms as a reference, determine its start-up time position, and select a segment of the waveform as a fixed window;

[0032] Step 5.2: Select a fixed range before and after the start time of another fault recording waveform as a sliding window. Within this window, starting from each sampling point, select a segment of waveform and compare it with the corresponding fixed window waveform from left to right.

[0033] Step 5.3: Calculate the Euclidean distance between the corresponding sampling points and select the minimum point as the alignment point for waveform consistency alignment.

[0034] Furthermore, the waveform similarity matching in step 6 includes the following steps:

[0035] Step 6.1: Shrink the fault waveform sequence in the fault waveform file data so that the coarse-grained waveform sequence replaces the original waveform sequence;

[0036] Step 6.2: Use the DTW algorithm to find the corresponding optimal bending path for the coarse-grained waveform sequence;

[0037] Step 6.3: Refine the obtained optimal bending path with the original waveform sequence size path and control the search range of the bending path to finally obtain the optimal bending path and minimum distance of the waveforms of the two fault recording files to be matched.

[0038] Furthermore, the method for shrinking the fault waveform sequence in the fault waveform file data is as follows: take the average value of multiple consecutive data points in the original waveform sequence as the data points of the coarse-grained waveform sequence.

[0039] The advantages and positive effects of this invention are:

[0040] This invention addresses the inconsistency in sampling frequencies between the protection device and the fault recorder by converting the sampling frequency of the extracted fault waveform data. Then, it uses Euclidean distance to align the waveforms of similar waveform files to achieve clock synchronization. Finally, the FDTW algorithm is used for intelligent comparison of the fault waveform data. This invention enables consistency processing of similar waveform files, achieving consistent sampling frequencies and clock synchronization, and realizing similarity matching between the waveforms recorded by the protection device and the waveforms recorded by the fault recorder in a vacuum semantic environment. Attached Figure Description

[0041] Figure 1 This is a flowchart of the present invention;

[0042] Figure 2 This is a schematic diagram of the fault recording waveform sampling frequency conversion of the present invention;

[0043] Figure 3 The process of finding the optimal alignment path for the DTW algorithm;

[0044] Figure 4 This is a waveform diagram of phase A voltage extracted from M and N in an embodiment of the present invention;

[0045] Figure 5 This is a diagram illustrating the calculation of the sliding window position according to an embodiment of the present invention;

[0046] Figure 6 This is a diagram showing the calculation of the sliding window position when the waveform to be matched is abnormal, according to an embodiment of the present invention.

[0047] Figure 7 This is a waveform diagram of a fault recording of the protection device without channel identification information according to an embodiment of the present invention.

[0048] Figure 8 This embodiment of the invention uses the minimum distance map obtained by the traditional DTW algorithm.

[0049] Figure 9 This is a minimum distance map using the FDTW algorithm in an embodiment of the present invention. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to the accompanying drawings.

[0051] Intelligent comparison method for fault recording data based on fast dynamic time warping algorithm, such as Figure 1 As shown, it includes the following steps:

[0052] Step 1: Extract the fault waveform data that needs to be compared.

[0053] Step 2: Determine whether the sampling frequencies of the fault recording data are consistent. If they are consistent, proceed to Step 4; otherwise, proceed to Step 3.

[0054] Step 3: Convert the sampling frequency of the fault recording data to make the sampling frequency of the fault recording data consistent.

[0055] When sampling the original analog signal, the Nyquist-Shannon sampling theorem must be satisfied, meaning the sampling frequency must be greater than or equal to twice the maximum frequency of the sampled signal to ensure complete recovery of the original analog signal's information. This invention employs resampling technology to transform the sampling frequency of the original analog signal, achieving consistency in sampling frequencies between the protection device and the fault recorder's source waveform files. Since the fault recorder's sampling frequency is higher than that of the relay protection device, to reduce the computational load during comparison while ensuring data comparison accuracy, it is necessary to convert the fault waveform with the higher sampling frequency, such as... Figure 2 As shown, this invention uses the Lagrange interpolation resampling method to convert the sampling frequency of fault recording data:

[0056] The original fault recording data waveform sampling rate f o Switch to the new sampling frequency f n Calculate the waveform length F of the fault recording data under the new sampling rate:

[0057]

[0058] Where G is the waveform length of the original fault recording data;

[0059] Calculate the specific location c of the interpolation within the sampling interval of the original fault recording data. k :

[0060]

[0061] Where k is the sample value index, and its range is 1≤k≤F; the time exponent c is set (1≤c≤G), and:

[0062] c≤c k ≤c+1

[0063] The window length is determined, where the highest order of the Lagrange interpolation polynomial is 2×v, the size of the window function is 2×v+1, and the time factors for the 2×v+1 original samples are as follows:

[0064] cv,…,c-1,c,c+1,…,c+v

[0065] The resampling weights are calculated for all points within the window length using the Lagrange resampling weight formula:

[0066]

[0067] Where i = -v, -v+1, ..., v are the window positions where the interpolation points are located, and l i Let $l$ be the resampling weight for the interpolation point $i$, and $j$ be the window position of the other interpolation points besides $i$, where $j = -v, -v+1, ..., v$ and $j ≠ i$. Based on the magnitude and resampling weight $l$ of each interpolation point... i Calculate the data amplitude at the interpolation point and obtain the weighted polynomial resampled sequence L(k) of 2×v+1 original sample values:

[0068]

[0069] Where m(ci) is the sampling sequence of the original fault recording data.

[0070] Step 4: Determine if the time synchronization of the fault recording data is consistent. If it is, proceed to step 6; otherwise, proceed to step 5.

[0071] Step 5: Align the waveforms of the fault recording data for consistency.

[0072] Due to differences in manufacturers, models, and years of operation of fault recording devices, even if two sets of recording devices have the same settings, there will be a difference in start-up time within the millisecond range (generally less than 20ms) when facing the same disturbance in the power grid. This results in inaccurate time synchronization between the two sets of fault recording devices, making it difficult to determine whether the recording file was triggered by the same disturbance in the power grid.

[0073] Euclidean distance (ED) measures the absolute distance between points in a multidimensional space. A smaller distance value indicates higher similarity. This method is computationally inefficient and time-saving. The expression for ED in n-dimensional space is:

[0074]

[0075] Where x and y are any two points in space, with coordinates x and y respectively. i and y i (i = 1, 2, ..., n), where d(x, y) is the Euclidean distance between x and y.

[0076] To align the start-up times of waveforms from the same fault source, this invention uses a sliding window to calculate Euclidean distances for point-by-point comparison, selecting the minimum value point as the optimal alignment point.

[0077] Step 5.1: Using one of the fault recording waveforms as a reference, determine its start-up time position, and select a segment of the waveform as a fixed window.

[0078] Step 5.2: Select a fixed range before and after the start time of another fault recording waveform as a sliding window. Within this window, starting from each sampling point, select a segment of waveform and compare it with the corresponding fixed window waveform from left to right.

[0079] Step 5.3: Calculate the Euclidean distance between the corresponding sampling points and select the minimum point as the alignment point for waveform consistency alignment.

[0080] Step 6: Perform waveform similarity matching on the fault recording data and compare and analyze the matching results.

[0081] Fault waveform recordings are acquired and transmitted in real time via devices such as current transformers and merging units, and their waveform data is a time series. Due to factors such as waveform discontinuities, missing waveforms, amplitude scaling or shifting, and noise interference, waveforms in fault waveform recordings from the same source may differ or be distorted, increasing the difficulty of waveform similarity matching. Therefore, it is necessary to find an effective time series similarity matching method. This invention uses the FDTW algorithm to perform similarity matching on fault waveform recordings.

[0082] The main improvements of FDTW over DTW are in constraints and abstraction. The former reduces the number of searches by limiting the search space of the optimal curved path, while the latter reduces the order of the curved matrix by shrinking the time series in order to find the optimal curved path. The found curved path is then mapped to the original curved matrix to find the minimum distance of the time series to be matched.

[0083] The main idea of ​​fault recording waveform similarity matching based on the FDTW algorithm is to shrink the waveform sequence data to be matched using constraint and abstraction methods, use the DTW algorithm to find the curvature path of the shrunken waveform sequence, restore the waveform sequence to its original size, and find the minimum distance between the waveform sequences to be matched. This method can effectively reduce the time complexity of the DTW algorithm while ensuring the accuracy of similarity matching. The matching process is as follows:

[0084] Step 6.1: Shrink the fault waveform sequence in the fault waveform file data so that the coarse-grained waveform sequence replaces the original waveform sequence. The shrinkage method is to take the average value of multiple consecutive data points in the original waveform sequence as the data points of the coarse-grained waveform sequence.

[0085] Step 6.2, as follows Figure 3 As shown, the optimal bending path is obtained by using the DTW algorithm on the coarse-grained waveform sequence.

[0086] Step 6.3: Refine the obtained optimal bending path with the original waveform sequence size path and control the search range of the bending path to finally obtain the optimal bending path and minimum distance of the waveforms of the two fault recording files to be matched.

[0087] Based on the above-mentioned intelligent comparison method for fault recording data based on the fast dynamic time warping algorithm, the effectiveness of the present invention is verified by matching the fault recording files of the same source at a certain moment in a smart substation.

[0088] The fault recording files M and N, extracted from a smart substation at a certain moment, represent a two-phase ground fault (AB phases). After the fault occurred, the three-phase circuit breaker was disconnected, and three-phase reclosing was employed. The sampling frequency of the fault recorder was 4000Hz, and the sampling period was 0.25 milliseconds.

[0089] Taking the phase A voltage waveform as an example, waveform consistency alignment calculation is performed. The phase A voltage waveform extracted from M and N is as follows: Figure 4 As shown. By Figure 4 It can be seen that there is a time difference of approximately 15 milliseconds between the fault recording files M and N at the start of recording.

[0090] Starting from the initial position of waveform M, 10 sampling points are taken as a fixed window. Using the initial position of waveform N as a reference point, 200 sampling points are taken to the left and right as sliding matching ranges. Each point within the matching range of waveform N is used as a starting point, and the waveform in the fixed window is compared point-by-point from left to right. The change in the position of the sliding window, calculated using the ED method, is shown in the figure below. Figure 5 As shown:

[0091] Depend on Figure 5 It can be seen that the window position of the minimum value of ED is at the 60th point to the left of the reference point, which is 60 × 0.25 = 15 milliseconds to the left of the start time of waveform N recording. This is the best alignment point of the recorded waveform N.

[0092] To verify that the ED method can accurately synchronize time even when the waveform to be matched is abnormal, the amplitude of waveform N was magnified by a factor of 1.1, while other parameters remained unchanged. The ED method was then used again to calculate the change in the sliding window position, as shown in the figure. Figure 6 .

[0093] Depend on Figure 6 It can be seen that when the waveform to be matched is abnormal, the window position of the minimum value of ED is still at the 60th point to the left of the reference point, which can achieve accurate time synchronization.

[0094] To verify the performance of the FDTW algorithm in fault recording waveform similarity matching, waveforms from recording files without channel identification information in a certain protection device were extracted and numbered sequentially, such as... Figure 7 As shown.

[0095] Using the waveforms of the corresponding 8 channels of the fault recorder as a benchmark, similarity matching was performed on the waveforms of the protection device recording files without channel identification information one by one based on the traditional DTW algorithm and the FDTW algorithm. The minimum distance diagram is shown below. Figure 8 and 9 As shown in Table 1, the waveform similarity matching results are as follows.

[0096] Table 1 Waveform similarity matching results

[0097]

[0098]

[0099] Depend on Figure 8 and 9As shown in Table 1, both the traditional DTW algorithm and the FDTW algorithm can accurately match waveforms from the same source and exhibit high stability. However, the FDTW algorithm has a significantly shorter matching time, and its computational efficiency is about 10 times that of the traditional DTW algorithm, resulting in better real-time matching performance.

[0100] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A method for intelligent comparison of fault recording data based on a fast dynamic time warping algorithm, characterized in that: Includes the following steps: Step 1: Extract the fault waveform data that needs to be compared; Step 2: Determine whether the sampling frequencies of the fault recording data are consistent. If they are consistent, proceed to Step 4; otherwise, proceed to Step 3. Step 3: Convert the sampling frequency of the fault recording data to make the sampling frequency of the fault recording data consistent. The specific implementation method of step 3 is as follows: The sampling frequency of the fault recording file data is converted using the Lagrange interpolation resampling method. The original fault recording data waveform sampling rate Switch to new sampling frequency Calculate the waveform length of the fault recording data at the new sampling frequency. : ； in, The length of the original fault recording data; Calculate the specific location of the interpolation within the sampling interval of the original fault recording data. : ； in, The sample value number is the range of the sample value. Set time index ,in, ,and: ； Determine the window length, where the highest order of the Lagrange interpolation polynomial is . The size of the window function is , The time factors for the original sample values ​​are as follows: ； The resampling weights are calculated for all points within the window length using the Lagrange resampling weight formula: ； in, The window position where the interpolation point is located. for The resampling weights at the interpolation points, To remove The window positions of other interpolation points outside the point, and Based on the magnitude and resampling weights at each interpolation point Calculate the data magnitude at the interpolation point and obtain the weighted values. A resampled sequence of polynomials of the original sample values : ； in, This is the sampling sequence of the original fault recording data; Step 4: Determine if the time synchronization of the fault recording data is consistent. If it is consistent, proceed to step 6; otherwise, proceed to step 5. Step 5: Align the waveforms of the fault recording data for consistency. Step 5 includes the following steps: Step 5.1: Using one of the fault recording waveforms as a reference, determine its start-up time position, and select a segment of the waveform as a fixed window; Step 5.2: Select a fixed range before and after the start time of another fault recording waveform as a sliding window. Within this window, starting from each sampling point, select a segment of waveform and compare it with the corresponding fixed window waveform from left to right. Step 5.3: Calculate the Euclidean distance between the corresponding sampling points, and select the minimum point as the alignment point for waveform consistency alignment; Step 6: Perform waveform similarity matching on the fault recording data and compare and analyze the matching results; Step 6, waveform similarity matching, includes the following steps: Step 6.1: Shrink the fault waveform sequence in the fault waveform file data so that the coarse-grained waveform sequence replaces the original waveform sequence; Step 6.2: Use the DTW algorithm to find the corresponding optimal bending path for the coarse-grained waveform sequence; Step 6.3: Refine the obtained optimal bending path with the original waveform sequence size path and control the search range of the bending path to finally obtain the optimal bending path and minimum distance of the waveforms of the two fault recording files to be matched.

2. The intelligent comparison method for fault recording data based on the fast dynamic time warping algorithm according to claim 1, characterized in that: The method for shrinking the fault waveform sequence in the fault waveform file data is as follows: take the average value of multiple consecutive data points in the original waveform sequence as the data points of the coarse-grained waveform sequence.

Images