A method and system for diagnosing faults of a flexible direct current capacitor
Patent Information
- Application Number
- CN202610894454.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-22
- Publication Date
- 2026-09-15
Smart Images
Figure CN122754652A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of condition monitoring technology for flexible DC transmission equipment, and specifically to a method and system for diagnosing faults in flexible DC capacitors. Background Technology
[0002] In flexible DC transmission systems, the DC support capacitor (usually a metallized film capacitor) is one of the core components of the converter valve, and its operating status directly affects the system's voltage stability and energy transmission reliability. As the capacitor's operating time increases, faults such as dielectric aging, deterioration of self-healing properties, and thermal breakdown gradually emerge. If not detected in time, these faults may lead to sudden capacitor failure, causing converter valve shutdown or even equipment damage. Therefore, online fault diagnosis of flexible DC capacitors is of great significance.
[0003] In existing technologies, time-frequency analysis methods based on wavelet packet decomposition are used to extract self-healing pulses and degradation characteristics of capacitors. Under strong harmonic backgrounds and non-stationary impact conditions, extracting weak self-healing pulses requires multi-level time-frequency decomposition, resulting in single-diagnosis delays of tens of milliseconds, far exceeding the microsecond to millisecond real-time protection requirements of flexible DC valve control systems. If the sampling rate is forced to be reduced or the number of decomposition levels decreased, early degradation signals in the high-frequency band are lost, causing fault diagnosis to occur later than the capacitor's thermal breakdown critical point, thus losing its early warning significance. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for diagnosing faults in flexible DC capacitors, so as to solve the problems mentioned above.
[0005] The objective of this invention can be achieved through the following technical solutions:
[0006] A method for diagnosing faults in a flexible DC capacitor includes the following steps:
[0007] Step 1: Obtain the voltage time-domain sequence and current time-domain sequence across the flexible DC capacitor, and merge the voltage time-domain sequence and current time-domain sequence into the initial observation sequence;
[0008] Step 2: Based on the preset phase segmentation basis, the initial observation sequence is divided into multiple phase subsequences. Parallel filtering operations are performed on each phase subsequence, and all the filtered phase subsequences are recombined into the first decomposed sequence.
[0009] Step 3: The first decomposition sequence is imported into the equivalent translocation cascade structure. The equivalent translocation cascade structure is decomposed layer by layer through filtering cascade after forward downsampling operation to generate a second decomposition sequence covering different frequency ranges.
[0010] Step 4: Divide the second decomposition sequence into a high-frequency pulse sub-band and a low-frequency degradation sub-band according to the frequency band. Extract the peak energy and pulse repetition interval from the high-frequency pulse sub-band as self-healing feature values, and calculate the energy entropy from the low-frequency degradation sub-band as degradation feature values.
[0011] Step 5: Jointly judge the self-healing characteristic value and the deterioration characteristic value. Based on the joint judgment result, output the comprehensive early warning index. When the comprehensive early warning index exceeds the preset alarm threshold, output the fault early warning signal.
[0012] As a further aspect of the present invention: the recombining of all filtered phase subsequences into a first decomposition sequence specifically includes:
[0013] The initial observation sequence is subjected to multi-phase delay broadening, and then each phase subsequence is obtained by equally interval extraction according to the phase segmentation base.
[0014] The preset filter coefficients are split into multiple coefficient subsets according to the same phase segmentation base, so that each coefficient subset corresponds one-to-one with a phase subsequence, and multiplication and addition operations are performed on each phase subsequence and its corresponding coefficient subset at the same time.
[0015] The results of all multiplication and addition operations are shifted and accumulated in reverse order of extraction, and then recombined into the first decomposition sequence.
[0016] As a further aspect of the present invention: the step of splitting the preset filter coefficients into multiple coefficient subsets according to the same phase segmentation basis specifically includes:
[0017] Extract all coefficients with a modulus phase segmentation base of zero from the preset filter coefficients as the first coefficient subset;
[0018] Extract coefficients with a modulus of one as the second coefficient subset, and so on until the coefficients with a modulus of one equal to the phase segmentation base minus one are selected as the last coefficient subset.
[0019] The coefficients within each subset of coefficients are stored consecutively in ascending order of their indices.
[0020] Each subset of coefficients is latched into the constant register of the corresponding parallel multiply-accumulate unit.
[0021] As a further aspect of the present invention: the generation of the second decomposition sequence covering different frequency ranges specifically includes:
[0022] The first decomposed sequence is simultaneously fed into multiple delay branches, and each branch performs downsampling extraction according to the phase segmentation basis to obtain multiple extracted subsequences;
[0023] Each extracted subsequence is convolved with the low-pass filter coefficient set and the high-pass filter coefficient set respectively to obtain the low-pass convolution result and the high-pass convolution result.
[0024] The low-pass convolution results of each branch are phase-recombined in reverse order of extraction to form the approximate decomposition sequence of the next layer. The high-pass convolution results of each branch are recombined in the same reverse order to form the detail decomposition sequence of the next layer.
[0025] Use the approximate decomposition sequence of the current layer as the input of the next layer, repeat the operation until the preset number of layers is reached, and merge the detailed decomposition sequences of all layers with the approximate decomposition sequence of the last layer into the second decomposition sequence.
[0026] As a further aspect of the present invention: the formation of the approximate decomposition sequence for the next layer specifically includes:
[0027] The length of the low-pass convolution result of each branch is counted. The longest branch length is used as the reference length. For branches that are shorter than the reference length, zeros are added to the end until the lengths are equal.
[0028] The low-pass convolution results of all branches are cross-arranged in reverse order of branch number, that is, the first element of the last branch is used as the first element of the recombined sequence, the first element of the second last branch is used as the second element of the recombined sequence, and so on, until the first element of all branches is arranged.
[0029] Arrange the subsequent elements of each branch according to the same rules until all elements are arranged to obtain the intermediate recombination sequence;
[0030] A smoothing operation is performed on the intermediate recombination sequence by averaging adjacent elements to form an approximate decomposition sequence for the next layer.
[0031] As a further aspect of the present invention: the process of constructing the self-healing feature value is as follows:
[0032] Traverse all sampling points within the high-frequency pulse sub-band, mark the points whose amplitude exceeds both the previous and next sampling points as pulse peak points, and record the amplitude and occurrence time of each pulse peak point;
[0033] Calculate the sum of squares of the amplitudes at all pulse peaks and use the sum of squares as the peak energy.
[0034] Calculate the difference between the times when adjacent pulse peaks occur, take the reciprocal of all differences and calculate the arithmetic mean, and use the average value as the pulse repetition interval;
[0035] The peak energy and the pulse repetition interval together constitute the self-healing characteristic value.
[0036] As a further aspect of the present invention: the calculation process of the degradation characteristic value is as follows:
[0037] The low-frequency degraded subband is divided into multiple consecutive time segments with a fixed duration. The root mean square value of the amplitude of all sampling points in each time segment is calculated to obtain the root mean square value sequence.
[0038] Divide the root mean square value sequence into equally spaced intervals, count the frequency of the root mean square value in each interval, and divide each frequency by the total number of segments to obtain the probability distribution.
[0039] Calculate the product of each probability value and its natural logarithm, sum the negatives of all products, and use the sum as the energy entropy;
[0040] The output energy entropy is used as a degradation characteristic value.
[0041] As a further aspect of the present invention: the joint discrimination specifically includes:
[0042] Extract the quotient of the pulse repetition interval divided by the peak energy from the self-healing feature value, and then calculate the reciprocal of the quotient as the first discriminant;
[0043] Extract the deterioration feature value and calculate its arithmetic square root as the second discriminant;
[0044] Multiply the first discriminant by the second discriminant, and then add the product to 1 / 10 of the deterioration eigenvalue to obtain the comprehensive early warning index;
[0045] The comprehensive early warning index is compared with the preset alarm threshold. When the comprehensive early warning index exceeds the preset alarm threshold, a fault early warning signal is output.
[0046] A fault diagnosis system for flexible DC capacitors includes:
[0047] The signal acquisition and merging module acquires the voltage time-domain sequence and current time-domain sequence across the flexible DC capacitor, and merges the voltage time-domain sequence and current time-domain sequence into an initial observation sequence.
[0048] The parallel filtering and recombination module divides the initial observation sequence into multiple phase subsequences according to the preset phase segmentation basis, performs parallel filtering operations on each phase subsequence, and recombines all the filtered phase subsequences into the first decomposed sequence.
[0049] The cascade decomposition module imports the first decomposition sequence into the equivalent translocation cascade structure. The equivalent translocation cascade structure decomposes the sequence layer by layer through filtering cascade after forward downsampling operation, generating a second decomposition sequence covering different frequency ranges.
[0050] The feature extraction module divides the second decomposition sequence into a high-frequency pulse sub-band and a low-frequency degradation sub-band according to the frequency band. It extracts the peak energy and pulse repetition interval from the high-frequency pulse sub-band as self-healing feature values and calculates the energy entropy from the low-frequency degradation sub-band as degradation feature values.
[0051] The joint discrimination and early warning module jointly discriminates the self-healing characteristic value and the deterioration characteristic value. Based on the joint discrimination result, it outputs a comprehensive early warning index. When the comprehensive early warning index exceeds the preset alarm threshold, it outputs a fault early warning signal.
[0052] The beneficial effects of this invention are:
[0053] (1) By adopting phase-segmented parallel filtering and equivalent translocation cascade structure, the single diagnostic delay of multi-layer time-frequency decomposition is compressed to the microsecond level without reducing the sampling rate or the number of decomposition layers, which meets the real-time requirements of the flexible DC valve control system for fault detection. At the same time, the weak self-healing pulse and early degradation signal in the high-frequency band are completely preserved, avoiding the loss of key features caused by the reduction of the sampling rate by traditional methods.
[0054] (2) By extracting the peak energy and pulse repetition interval of the high-frequency pulse sub-band as self-healing characteristic value and the energy entropy of the low-frequency deterioration sub-band as deterioration characteristic value, and using the joint discrimination rule to calculate the comprehensive early warning index, it is possible to effectively distinguish between normal self-healing events of capacitors and continuous dielectric deterioration processes, reduce false alarms caused by frequent self-healing events, improve the accuracy of fault diagnosis results, and provide more reliable early warning basis for maintenance personnel. Attached Figure Description
[0055] The invention will now be further described with reference to the accompanying drawings.
[0056] Figure 1 This is a flowchart of the method of the present invention;
[0057] Figure 2 This is a system block diagram of the present invention. Detailed Implementation
[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0059] Please see Figure 1 As shown, the present invention provides a fault diagnosis method for flexible DC capacitors, comprising the following steps:
[0060] Step 1: Obtain the voltage time-domain sequence and current time-domain sequence across the flexible DC capacitor, and merge the voltage time-domain sequence and current time-domain sequence into the initial observation sequence;
[0061] Step 2: Based on the preset phase segmentation basis, the initial observation sequence is divided into multiple phase subsequences. Parallel filtering operations are performed on each phase subsequence, and all the filtered phase subsequences are recombined into the first decomposed sequence.
[0062] Step 3: The first decomposition sequence is imported into the equivalent translocation cascade structure. The equivalent translocation cascade structure is decomposed layer by layer through filtering cascade after forward downsampling operation to generate a second decomposition sequence covering different frequency ranges.
[0063] Step 4: Divide the second decomposition sequence into a high-frequency pulse sub-band and a low-frequency degradation sub-band according to the frequency band. Extract the peak energy and pulse repetition interval from the high-frequency pulse sub-band as self-healing feature values, and calculate the energy entropy from the low-frequency degradation sub-band as degradation feature values.
[0064] Step 5: Jointly judge the self-healing characteristic value and the deterioration characteristic value. Based on the joint judgment result, output the comprehensive early warning index. When the comprehensive early warning index exceeds the preset alarm threshold, output the fault early warning signal.
[0065] In step one, the voltage time-domain sequence and current time-domain sequence across the flexible DC capacitor are obtained, and the voltage time-domain sequence and current time-domain sequence are merged into an initial observation sequence, specifically including:
[0066] A resistive voltage divider type voltage sensor is installed in parallel across the two ends of a flexible DC capacitor, and a through-hole current transformer is connected in series in the grounding branch of the capacitor. The output signals of the two sensors are respectively connected to an analog-to-digital converter with synchronous sample-and-hold function. The sampling frequency is set to 1 MHz, the sampling resolution is 16 bits, and the continuous acquisition time is 1 ms to obtain voltage time-domain sequence and current time-domain sequence. Each sequence contains 1024 discrete sampling points, and the sampling time is strictly aligned. The voltage time-domain sequence is denoted as U[0] to U
[1023] , and the current time-domain sequence is denoted as I[0] to I
[1023] . All values are expressed in volts and amperes in the International System of Units (SI).
[0067] The voltage time-domain sequence and the current time-domain sequence are merged according to the correspondence of the same sampling time. The specific merging method is as follows: for the k-th sampling point (k ranges from 0 to 1023), the voltage value U[k] is written first, followed by the current value I[k], forming a pair of data. After processing all 1024 sampling points in sequence, an alternating sequence of length 2048 is obtained. This sequence is the initial observation sequence, and its element order is U[0], I[0], U[1], I[1], …, U
[1023] , I
[1023] .
[0068] To facilitate subsequent processing, the initial observation sequence is stored in contiguous storage units, with each data point occupying 2 bytes. If any sampling point data is abnormal during the acquisition process (e.g., exceeding the preset range or experiencing excessively large abrupt changes), the current 1-millisecond data window is discarded and re-acquisition is performed until a complete and valid initial observation sequence is obtained. All acquisition and merging operations are completed online under normal operating conditions of the flexible DC capacitor, without altering the electrical connection between the capacitor and the converter valve.
[0069] The initial observation sequence obtained in the above manner simultaneously contains dynamic information about the voltage across the capacitor and dynamic information about the current flowing through the capacitor, with a one-to-one correspondence between voltage and current in the time domain. This provides an aligned data foundation for subsequent phase segmentation and parallel filtering. The length of this sequence and the sampling parameters can be adjusted according to the actual diagnostic accuracy and computational resources, but the merging rule always maintains the alternating order of voltage first and current last.
[0070] In step two, the initial observation sequence is divided into multiple phase subsequences according to a preset phase segmentation basis. Parallel filtering is performed on each phase subsequence, and all filtered phase subsequences are recombined into the first decomposed sequence. Specifically, this includes:
[0071] The phase segmentation base is set to 4. A multi-phase delay broadening operation is performed on the initial observation sequence obtained in step one. Specifically, the initial observation sequence is copied four times and delayed by 0, 1, 2, and 3 sampling periods respectively, resulting in four delayed sequences. Then, each delayed sequence is sampled at equal intervals according to the phase segmentation base of 4. That is, starting from the beginning of each sequence, one data point is taken every 4 sampling points until all sampling points are taken, resulting in four phase subsequences, denoted as phase subsequence 0, phase subsequence 1, phase subsequence 2, and phase subsequence 3. The length of each phase subsequence is one-quarter of the original sequence length (rounded up).
[0072] Prepare a set of preset filter coefficients, which are low-pass filter coefficients, a total of 16, numbered sequentially from 0 to 15: 0.0002, 0.0015, 0.0090, 0.0305, 0.0706, 0.1211, 0.1609, 0.1724, 0.1609, 0.1211, 0.0706, 0.0305, 0.0090, 0.0015, 0.0002, 0.0000. The preset filter coefficients are split into four subsets using the same phase division base of 4. The splitting rules are as follows: the coefficients whose index modulo 4 equals 0 are extracted as the first subset, i.e., the coefficients corresponding to indices 0, 4, 8, and 12, with values of 0.0002, 0.0706, 0.1609, and 0.0090 respectively, and stored consecutively in ascending order of index; the coefficients whose modulo 4 equals 1 are extracted as the second subset, i.e., the coefficients corresponding to indices 1, 5, 9, and 13, with values of 0. 0015, 0.1211, 0.1211, 0.0015; extract the coefficients modulo 4 equal to 2 as the third coefficient subset, i.e., the coefficients corresponding to indices 2, 6, 10, and 14, with values of 0.0090, 0.1609, 0.0706, and 0.0002 respectively; extract the coefficients modulo 4 equal to 3 as the fourth coefficient subset, i.e., the coefficients corresponding to indices 3, 7, 11, and 15, with values of 0.0305, 0.1724, 0.0305, and 0.0000 respectively. After storing the coefficients in each coefficient subset consecutively in ascending index order, they are latched into the constant registers of the corresponding parallel multiply-accumulate units, with the first coefficient subset latched into the 0th multiply-accumulate unit, the second coefficient subset into the 1st multiply-accumulate unit, and so on.
[0073] The four phase subsequences are fed into their respective parallel multiply-accumulate units. Each multiply-accumulate unit performs a multiplication operation on each data point in the received phase subsequence with each coefficient in the corresponding coefficient subset, and sums all the multiplications to obtain the convolution result of that phase subsequence. Specifically, for the 0th phase subsequence, its first data point is multiplied by each of the four coefficients (0.0002, 0.0706, 0.1609, 0.0090) in the first coefficient subset, and the sum is obtained to get the first convolution output value. Then, the phase subsequence is slid one position, and the above multiply-accumulate operation is repeated until all data has been calculated to obtain the convolution result sequence of that branch. The other branches are processed in the same way.
[0074] The convolutional result sequences from all four branches are shifted and accumulated in reverse order of extraction to reassemble into the first decomposition sequence. The specific operation of extraction in reverse order is as follows: First, find the maximum length of all branch convolutional result sequences, and pad the ends of branches shorter than this length with zeros; then, arrange the branches in reverse order by cross-referencing, i.e., the first element of the 3rd branch becomes the first element of the reassembled sequence, the first element of the 2nd branch becomes the second element, the first element of the 1st branch becomes the third element, and the first element of the 0th branch becomes the fourth element; then, arrange the second and third elements of each branch in the same order until all elements are arranged, obtaining an intermediate sequence. Each element of this intermediate sequence is added to its preceding element (if any) and divided by 2, i.e., the average of adjacent elements is calculated to complete the smoothing operation, and the output is the first decomposition sequence. The length of this first decomposition sequence is the same as the original initial observation sequence, but it has undergone equivalent processing of low-pass filtering and downsampling.
[0075] In step three, the first decomposition sequence is imported into the equivalent translocation cascade structure. The equivalent translocation cascade structure decomposes the sequence layer by layer through filtering cascades after forward downsampling operations, generating a second decomposition sequence covering different frequency ranges, specifically including:
[0076] The decomposition preset number of layers is set to 6, and the phase segmentation base is kept at 4. The first decomposed sequence obtained in step two is simultaneously fed into 4 delay branches. Each branch first performs multi-phase delay broadening, that is, branch 0 is not delayed, branch 1 is delayed by 1 sampling period, branch 2 is delayed by 2 sampling periods, and branch 3 is delayed by 3 sampling periods. Then, each branch performs downsampling at the phase segmentation base of 4, that is, starting from the beginning position of the delayed sequence of each branch, one data point is taken every 4 sampling points to obtain 4 extracted subsequences, which are denoted as the 0th extracted subsequence, the 1st extracted subsequence, the 2nd extracted subsequence, and the 3rd extracted subsequence. The length of each extracted subsequence is approximately one-quarter of the length of the first decomposed sequence.
[0077] Prepare two sets of preset filter coefficients: The low-pass filter coefficient set is the same as the preset filter coefficients in step two, with a total of 16 coefficients. The values of indices 0 to 15 are 0.0002, 0.0015, 0.0090, 0.0305, 0.0706, 0.1211, 0.1609, 0.1724, 0.1609, 0.1211, 0.0706, 0.0305, 0.0090, 0.0015, 0.0002, and 0.0000. The high-pass filter coefficient set consists of the corresponding orthogonal mirror filter coefficients, totaling 16 coefficients. The values from index 0 to 15 are 0.0000, -0.0002, -0.0015, -0.0090, -0.0305, -0.0706, -0.1211, -0.1609, 0.1724, -0.1609, 0.1211, -0.0706, 0.0305, -0.0090, 0.0015, and -0.0002, respectively. These two sets of coefficients are then split into four subsets based on a phase splitting radix of 4. The splitting rule is the same as in step two: coefficients with index numbers modulo 4 equal to 0, 1, 2, and 3 are extracted to form subsets. Each subset contains four coefficients, stored consecutively in ascending index order. The four subsets of the low-pass coefficients are latched into the low normal number registers of the four parallel multiply-accumulate units, and the four subsets of the high-pass coefficients are latched into the high normal number registers of the same set of multiply-accumulate units.
[0078] The four extracted sub-sequences are fed into their respective parallel multiply-accumulate units. Each unit first performs a multiply-accumulate operation with the low-pass coefficient subset to obtain the low-pass convolution result of that branch; then it performs a multiply-accumulate operation with the high-pass coefficient subset to obtain the high-pass convolution result of that branch. Specifically, for the low-pass convolution of branch 0, the first data point in the extracted sub-sequence of that branch is multiplied by the four coefficients (0.0002, 0.0706, 0.1609, 0.0090) in the low-pass coefficient subset, and the results are summed to obtain the first low-pass convolution output value. Then, the multiply-accumulate operation is repeated by sliding one data position until all data points have been calculated, obtaining the low-pass convolution result sequence for that branch. The calculation method for the high-pass convolution result sequence is the same for other branches, only the coefficient subsets differ.
[0079] For the low-pass convolution results, phase recombination is performed to form the approximate decomposition sequence for the next layer. The specific steps are as follows: First, the lengths of the low-pass convolution result sequences of the four branches are counted, and the maximum value is taken as the baseline length. Branches shorter than the baseline length are padded with zeros at the end to ensure all branches have equal length. Then, the low-pass convolution results of all branches are cross-arranged in reverse branch number order: the first element of the 3rd branch is used as the first element of the recombined sequence, the first element of the 2nd branch as the second element, the first element of the 1st branch as the third element, and the first element of the 0th branch as the fourth element. Next, the second element of each branch is arranged, i.e., the second element of the 3rd branch is used as the fifth element of the recombined sequence, and so on, until all elements of all branches are arranged, obtaining an intermediate recombined sequence. Finally, the average operation is performed on every two adjacent elements of this intermediate recombined sequence: the second element is added to the first element and divided by 2 to obtain the new second element; the third element is added to the second element (original value) and divided by 2 to obtain the new third element, and so on. The first element at the boundary remains unchanged. After smoothing, the approximate decomposition sequence for the next layer is obtained. The high-pass convolution results are processed in exactly the same way to obtain the decomposition-of-details sequence for the next layer.
[0080] The approximate decomposition sequence obtained from the current layer is used as the input for the next layer. All the above operations (i.e., delay, downsampling, convolution with low-pass and high-pass coefficients, recombination, and smoothing) are repeated until the preset 6-layer decomposition is completed. Each decomposition layer generates a detail decomposition sequence, resulting in a total of 6 detail decomposition sequences. After completing 6 layers, a final approximate decomposition sequence is also obtained. These 6 detail decomposition sequences and the final approximate decomposition sequence are concatenated in descending order of decomposition layer number (i.e., first layer detail first, second layer detail next, up to sixth layer detail, and finally the final approximate sequence) to form a total sequence. This total sequence is the second decomposition sequence covering different frequency ranges.
[0081] In step four, the second decomposition sequence is divided into a high-frequency pulse sub-band and a low-frequency degradation sub-band according to frequency band. Peak energy and pulse repetition interval are extracted from the high-frequency pulse sub-band as self-healing characteristic values, and energy entropy is calculated from the low-frequency degradation sub-band as degradation characteristic values. Specifically, this includes:
[0082] In step four, the second decomposition sequence obtained in step three is divided into subbands according to frequency range. The second decomposition sequence is composed of six detailed decomposition sequences and one final-level approximate decomposition sequence, spliced layer by layer. The first three detailed decomposition sequences correspond to the high-frequency range (center frequencies of 250 kHz, 125 kHz, and 62.5 kHz, respectively), the last three detailed decomposition sequences correspond to the mid-frequency range (center frequencies of 31.25 kHz, 15.625 kHz, and 7.8125 kHz, respectively), and the final-level approximate decomposition sequence corresponds to the low-frequency range (0 to 3.90625 kHz). The first three detailed decomposition sequences are merged to obtain the high-frequency pulse subband, which mainly contains transient pulse signals generated by capacitor self-healing events; the final-level approximate decomposition sequence is used as the low-frequency degradation subband, which mainly contains energy dissipation characteristics caused by capacitor dielectric degradation.
[0083] For high-frequency pulse sub-bands, peak detection is used to extract self-healing feature values. Each sampling point within the high-frequency pulse sub-band is traversed, comparing the amplitude of the current sampling point with the amplitudes of the previous and subsequent sampling points. If the amplitude of the current sampling point is strictly greater than both the amplitudes of the previous and subsequent sampling points, this sampling point is marked as a pulse peak point, and its amplitude and occurrence time in the entire sequence are recorded (represented by the sampling sequence number). After traversal, the amplitude sequence and time sequence of all pulse peak points are obtained. The sum of squares of the amplitudes of all pulse peak points is calculated; the result is the peak energy, which reflects the total energy released by the self-healing event. Let there be K pulse peak points, with their occurrence times as follows: , , …, (Unit: sampling period), then the time difference between adjacent pulse peak points is: ,in Take values from 1 to K-1. Calculate each difference. The reciprocal of the product is obtained. Then, calculate the arithmetic mean of these reciprocals. This mean is the pulse repetition interval, which can be mathematically expressed as: ;
[0084] Where R represents the pulse repetition interval, in Hertz; K is the total number of pulse peaks; For the first The occurrence time of each pulse peak point is dimensionless (measured in terms of the sampling period); the subtraction in the denominator represents the difference between adjacent moments. When K is less than 2, the pulse repetition interval is directly assigned to 0. The calculated peak energy (denoted as...) Together with the pulse repetition interval (denoted as R), the self-healing feature value is formed, which is stored in the form of a two-dimensional vector.
[0085] For low-frequency degraded subbands, the degradation characteristic value is calculated using the information entropy method.
[0086] The low-frequency degradation subband is divided into multiple consecutive time segments of fixed duration, with each time segment having a length of 0.5 milliseconds, corresponding to 512 sampling points (at a sampling rate of 1 MHz). If the total number of sampling points in the low-frequency degradation subband is N, then the number of time segments obtained is... , The mathematical operator represents rounding down, discarding the last segment with less than 512 points. For each time segment, the root mean square (RMS) value of the amplitudes of all sampling points within it is calculated. Specifically, the sum of the squares of the amplitudes of each sampling point within the segment is divided by the total number of sampling points in the segment (512), and the square root is taken to obtain the RMS value of that segment. After traversing all M segments, a sequence of RMS values of length M is obtained. Then, the RMS value sequence is divided into equally spaced intervals, dividing the range of RMS values from the minimum to the maximum value into 32 intervals. The width of each interval is (maximum value minus minimum value) divided by 32. The frequency of the RMS value in each interval is counted, i.e., the number of segments where the RMS value falls within that interval, resulting in a frequency sequence. , …, Divide each frequency by the total number of segments M to obtain the probability value for each interval. , where j ranges from 1 to 32. Next, the product of each probability value and its natural logarithm is calculated. For the case where the probability is 0, we define 0 × ln0 = 0. Finally, by summing the negatives of all products, the result is the energy entropy, whose mathematical expression is: ;
[0087] in, Represents energy entropy, which is dimensionless; For the first The probability values for each interval range from 0 to 1; ln represents the natural logarithm; the summation symbol represents the accumulation of all terms across the 32 intervals. H reaches its maximum value when all probability values are equal; H approaches 0 when the probability distribution is highly uneven. This energy entropy reflects the degree of disorder in the energy distribution within the low-frequency degradation sub-band: when the capacitor is operating normally, the low-frequency energy distribution is relatively uniform, and the entropy value is high; when dielectric degradation intensifies, energy concentrates in certain specific frequency bands or time segments, leading to a decrease in the entropy value. The calculated energy entropy is directly output as the degradation characteristic value, denoted as D.
[0088] Through the above process, self-healing characteristic values (including peak energy) are obtained from the high-frequency pulse subbands. The degradation characteristic value D is obtained from the low-frequency degradation subband (and pulse repetition interval R), providing a clear numerical basis for the joint discrimination in step five.
[0089] In step five, the self-healing characteristic value and the degradation characteristic value are jointly judged. Based on the joint judgment result, a comprehensive early warning index is output. When the comprehensive early warning index exceeds the preset alarm threshold, a fault early warning signal is output, specifically including:
[0090] In step five, two components, pulse repetition interval and peak energy, are extracted from the self-healing feature values obtained in step four. The pulse repetition interval is divided by the peak energy to obtain the quotient. Then, the reciprocal of this quotient is calculated, i.e., 1 is divided by the quotient. The result is used as the first discriminant. The physical meaning of the first discriminant is: when self-healing events are frequent (large pulse repetition interval) and the energy is small, this value is large; conversely, when self-healing events are rare or the energy is large, this value is small.
[0091] Extract the degradation feature value obtained in step four, i.e., the energy entropy. Calculate the arithmetic square root of this energy entropy by multiplying the degradation feature value by itself and then taking the square root, or by directly using the square root operation. The result is used as the second discriminant. Since the energy entropy is high (uniformly distributed) during normal operation and low during degradation, its arithmetic square root further narrows the numerical range, making the discrimination more sensitive.
[0092] Multiply the first and second discriminants to obtain a product. Then extract one-tenth of the degradation characteristic value (i.e., multiply the degradation characteristic value by 0.1), and add this one-tenth value to the product to obtain the comprehensive warning index. For example, if the degradation characteristic value is 0.8, then one-tenth of it is 0.08; if the product is 0.6, then the comprehensive warning index is 0.68. This comprehensive warning index comprehensively reflects the level of self-healing activity and the degree of degradation: when self-healing is frequent and degradation is severe, the index is high; when self-healing is scarce and the capacitor is healthy, the index is low.
[0093] The calculated comprehensive early warning index is compared with a preset alarm threshold. The preset alarm threshold is determined by experimental calibration: under laboratory conditions, an accelerated aging voltage is applied to a new capacitor of the same model, and the average comprehensive early warning index is recorded at the beginning of its lifespan; when the capacitor capacitance drops to 95% of its initial value, the comprehensive early warning index at this point is measured; 80% of this index is taken as the preset alarm threshold. After multiple tests, the preset alarm threshold in this embodiment is set to 0.65. When the real-time calculated comprehensive early warning index exceeds 0.65, a fault early warning signal is output, indicating that the flexible DC capacitor has entered the early fault stage and needs to be repaired or replaced. If the comprehensive early warning index does not exceed the threshold, monitoring continues, and no early warning signal is output. The entire joint discrimination process requires no manual intervention and is completed automatically based on real-time characteristic values.
[0094] Please see Figure 2 As shown, a fault diagnosis system for flexible DC capacitors includes:
[0095] The signal acquisition and merging module acquires the voltage time-domain sequence and current time-domain sequence across the flexible DC capacitor, and merges the voltage time-domain sequence and current time-domain sequence into an initial observation sequence.
[0096] The parallel filtering and recombination module divides the initial observation sequence into multiple phase subsequences according to the preset phase segmentation basis, performs parallel filtering operations on each phase subsequence, and recombines all the filtered phase subsequences into the first decomposed sequence.
[0097] The cascade decomposition module imports the first decomposition sequence into the equivalent translocation cascade structure. The equivalent translocation cascade structure decomposes the sequence layer by layer through filtering cascade after forward downsampling operation, generating a second decomposition sequence covering different frequency ranges.
[0098] The feature extraction module divides the second decomposition sequence into a high-frequency pulse sub-band and a low-frequency degradation sub-band according to the frequency band. It extracts the peak energy and pulse repetition interval from the high-frequency pulse sub-band as self-healing feature values and calculates the energy entropy from the low-frequency degradation sub-band as degradation feature values.
[0099] The joint discrimination and early warning module jointly discriminates the self-healing characteristic value and the deterioration characteristic value. Based on the joint discrimination result, it outputs a comprehensive early warning index. When the comprehensive early warning index exceeds the preset alarm threshold, it outputs a fault early warning signal.
[0100] The working principle of this invention is as follows: The voltage across the capacitor and the current on the ground side are collected and merged into an initial observation sequence. This sequence is then divided into multiple phase sub-sequences using a preset phase segmentation base. Each sub-sequence is filtered in parallel and recombined into a first decomposed sequence. This sequence is then imported into an equivalent translocation cascade structure, and through a forward downsampling operation followed by layer-by-layer filtering cascade decomposition, a second decomposed sequence covering different frequency ranges is generated. The second decomposed sequence is divided into a high-frequency pulse sub-band and a low-frequency degradation sub-band. Peak energy and pulse repetition interval are extracted from the high-frequency sub-band as self-healing characteristic values, and energy entropy is calculated from the low-frequency sub-band as degradation characteristic values. Finally, the self-healing characteristic values and degradation characteristic values are jointly judged to calculate a comprehensive early warning index. When the index exceeds a preset alarm threshold, a fault early warning signal is output.
[0101] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.
Claims
1. A method for diagnosing faults in a flexible DC capacitor, characterized in that, Includes the following steps: Step 1: Obtain the voltage time-domain sequence and current time-domain sequence across the flexible DC capacitor, and merge the voltage time-domain sequence and current time-domain sequence into the initial observation sequence; Step 2: Based on the preset phase segmentation basis, the initial observation sequence is divided into multiple phase subsequences. Parallel filtering operations are performed on each phase subsequence, and all the filtered phase subsequences are recombined into the first decomposed sequence. Step 3: The first decomposition sequence is imported into the equivalent translocation cascade structure. The equivalent translocation cascade structure is decomposed layer by layer through filtering cascade after forward downsampling operation to generate a second decomposition sequence covering different frequency ranges. Step 4: Divide the second decomposition sequence into a high-frequency pulse sub-band and a low-frequency degradation sub-band according to the frequency band. Extract the peak energy and pulse repetition interval from the high-frequency pulse sub-band as self-healing feature values, and calculate the energy entropy from the low-frequency degradation sub-band as degradation feature values. Step 5: Jointly judge the self-healing characteristic value and the deterioration characteristic value. Based on the joint judgment result, output the comprehensive early warning index. When the comprehensive early warning index exceeds the preset alarm threshold, output the fault early warning signal.
2. The method for fault diagnosis of a flexible DC capacitor according to claim 1, characterized in that, The step of recombining all filtered phase subsequences into the first decomposition sequence specifically includes: The initial observation sequence is subjected to multi-phase delay broadening, and then each phase subsequence is obtained by equally interval extraction according to the phase segmentation base. The preset filter coefficients are split into multiple coefficient subsets according to the same phase segmentation base, so that each coefficient subset corresponds one-to-one with a phase subsequence, and multiplication and addition operations are performed on each phase subsequence and its corresponding coefficient subset at the same time. The results of all multiplication and addition operations are shifted and accumulated in reverse order of extraction, and then recombined into the first decomposition sequence.
3. The method for fault diagnosis of a flexible DC capacitor according to claim 2, characterized in that, The step of splitting the preset filter coefficients into multiple coefficient subsets according to the same phase segmentation basis specifically includes: Extract all coefficients with a modulus phase segmentation base of zero from the preset filter coefficients as the first coefficient subset; Extract coefficients with a modulus of one as the second coefficient subset, and so on until the coefficients with a modulus of one equal to the phase segmentation base minus one are selected as the last coefficient subset. The coefficients within each subset of coefficients are stored consecutively in ascending order of their indices. Each subset of coefficients is latched into the constant register of the corresponding parallel multiply-accumulate unit.
4. The method for fault diagnosis of a flexible DC capacitor according to claim 1, characterized in that, The generation of the second decomposition sequence covering different frequency ranges specifically includes: The first decomposed sequence is simultaneously fed into multiple delay branches, and each branch performs downsampling extraction according to the phase segmentation basis to obtain multiple extracted subsequences; Each extracted subsequence is convolved with the low-pass filter coefficient set and the high-pass filter coefficient set respectively to obtain the low-pass convolution result and the high-pass convolution result. The low-pass convolution results of each branch are phase-recombined in reverse order of extraction to form the approximate decomposition sequence of the next layer. The high-pass convolution results of each branch are recombined in the same reverse order to form the detail decomposition sequence of the next layer. Use the approximate decomposition sequence of the current layer as the input of the next layer, repeat the operation until the preset number of layers is reached, and merge the detailed decomposition sequences of all layers with the approximate decomposition sequence of the last layer into the second decomposition sequence.
5. The method for fault diagnosis of a flexible DC capacitor according to claim 4, characterized in that, The approximate decomposition sequence for forming the next layer specifically includes: The length of the low-pass convolution result of each branch is counted. The longest branch length is used as the reference length. For branches that are shorter than the reference length, zeros are added to the end until the lengths are equal. The low-pass convolution results of all branches are cross-arranged in reverse order of branch number, that is, the first element of the last branch is used as the first element of the recombined sequence, the first element of the second last branch is used as the second element of the recombined sequence, and so on, until the first element of all branches is arranged. Arrange the subsequent elements of each branch according to the same rules until all elements are arranged to obtain the intermediate recombination sequence; A smoothing operation is performed on the intermediate recombination sequence by averaging adjacent elements to form an approximate decomposition sequence for the next layer.
6. The method for fault diagnosis of a flexible DC capacitor according to claim 1, characterized in that, The process of constructing the self-healing feature value is as follows: Traverse all sampling points within the high-frequency pulse sub-band, mark the points whose amplitude exceeds both the previous and next sampling points as pulse peak points, and record the amplitude and occurrence time of each pulse peak point; Calculate the sum of squares of the amplitudes at all pulse peaks and use the sum of squares as the peak energy. Calculate the difference between the times when adjacent pulse peaks occur, take the reciprocal of all differences and calculate the arithmetic mean, and use the average value as the pulse repetition interval; The peak energy and the pulse repetition interval together constitute the self-healing characteristic value.
7. The method for fault diagnosis of a flexible DC capacitor according to claim 1, characterized in that, The calculation process for the degradation characteristic value is as follows: The low-frequency degraded subband is divided into multiple consecutive time segments with a fixed duration. The root mean square value of the amplitude of all sampling points in each time segment is calculated to obtain the root mean square value sequence. Divide the root mean square value sequence into equally spaced intervals, count the frequency of the root mean square value in each interval, and divide each frequency by the total number of segments to obtain the probability distribution. Calculate the product of each probability value and its natural logarithm, sum the negatives of all products, and use the sum as the energy entropy; The output energy entropy is used as a degradation characteristic value.
8. The method for fault diagnosis of a flexible DC capacitor according to claim 1, characterized in that, The joint discrimination specifically includes: Extract the quotient of the pulse repetition interval divided by the peak energy from the self-healing feature value, and then calculate the reciprocal of the quotient as the first discriminant; Extract the deterioration feature value and calculate its arithmetic square root as the second discriminant; Multiply the first discriminant by the second discriminant, and then add the product to 1 / 10 of the deterioration eigenvalue to obtain the comprehensive early warning index; The comprehensive early warning index is compared with the preset alarm threshold. When the comprehensive early warning index exceeds the preset alarm threshold, a fault early warning signal is output.
9. A fault diagnosis system for flexible DC capacitors, characterized in that, A method for diagnosing faults in a flexible DC capacitor as described in any one of claims 1-8 includes: The signal acquisition and merging module acquires the voltage time-domain sequence and current time-domain sequence across the flexible DC capacitor, and merges the voltage time-domain sequence and current time-domain sequence into an initial observation sequence. The parallel filtering and recombination module divides the initial observation sequence into multiple phase subsequences according to the preset phase segmentation basis, performs parallel filtering operations on each phase subsequence, and recombines all the filtered phase subsequences into the first decomposed sequence. The cascade decomposition module imports the first decomposition sequence into the equivalent translocation cascade structure. The equivalent translocation cascade structure decomposes the sequence layer by layer through filtering cascade after forward downsampling operation, generating a second decomposition sequence covering different frequency ranges. The feature extraction module divides the second decomposition sequence into a high-frequency pulse sub-band and a low-frequency degradation sub-band according to the frequency band. It extracts the peak energy and pulse repetition interval from the high-frequency pulse sub-band as self-healing feature values and calculates the energy entropy from the low-frequency degradation sub-band as degradation feature values. The joint discrimination and early warning module jointly discriminates the self-healing characteristic value and the deterioration characteristic value. Based on the joint discrimination result, it outputs a comprehensive early warning index. When the comprehensive early warning index exceeds the preset alarm threshold, it outputs a fault early warning signal.