A power unit dynamic parameter optimization operation and maintenance management system of a power system

Abstract

The application relates to the technical field of power unit operation and maintenance management of a power system, and particularly discloses a power unit dynamic parameter optimization operation and maintenance management system of a power system, which collects complete transient time sequence data of gate potential, main loop current and terminal voltage in a switching transition process of a power unit, divides the data into five characteristic time zones according to first-order change rate mutation points, extracts local slope change rate, second-order derivative zero point density and short-time energy entropy of each time zone to combine into a high-dimensional morphological feature group, performs non-negative matrix blind source separation of mutual information minimization on the feature groups of continuous multiple switching cycles, separates out independent source component time sequence sequences of threshold value drift, conduction increment and thermal resistance change, and obtains each degradation amount through local fluctuation ratio adaptive sliding median filtering and exponential weighted trend locking, and compares the degradation amount with a preset threshold value to output an operation and maintenance prompt signal; the application can effectively separate degradation sources under the condition of parameter strong coupling, and improves the accuracy of operation and maintenance early warning.

Description

Technical Field

[0001] This invention relates to the field of power system power unit operation and maintenance management technology, and specifically to a power system power unit dynamic parameter optimization operation and maintenance management system. Background Technology

[0002] Currently, power systems widely utilize power cells, such as insulated-gate bipolar transistors (IGBTs) and silicon carbide metal-oxide-semiconductor field-effect transistors (SiC MOSFETs), in flexible DC converter valves, energy storage converters, and wind power converters. With increasing operating time, these power cells exhibit latent degradation phenomena, including threshold voltage drift, increased on-resistance, and changes in junction-case thermal impedance, due to hot carrier injection, gate oxide degradation, and accumulated electrothermal stress. To ensure system reliability, existing technologies typically collect gate potential, main circuit current, and terminal voltage data during the power cell switching process. These data, combined with recursive algorithms such as recursive least squares or Kalman filtering, are used to estimate these degradation parameters online and adjust control parameters such as switching frequency and dead time accordingly, or trigger maintenance warnings.

[0003] The existing technology has the following shortcomings: In the middle stage of power cell aging, under the condition that the threshold voltage drift, the increase in on-resistance and the change in junction-shell thermal impedance are strongly nonlinearly coupled and there is a lack of independent observation channels, how can we blindly separate the independent degradation source components from the single switching transient waveform without relying on an accurate physical model, so as to avoid the accumulation of decoupling errors caused by parameter unobservability of traditional recursive least squares or Kalman filtering over aging time and cause misjudgment of lifetime prediction model? Summary of the Invention

[0004] The purpose of this invention is to provide a power unit dynamic parameter optimization operation and maintenance management system for power systems to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A power unit dynamic parameter optimization operation and maintenance management system for a power system, comprising:

[0007] The transient data acquisition module collects complete transient timing data of the gate potential, main circuit current and terminal voltage of the power unit during the switching transition. The complete transient timing data covers the entire time period from before the gate drive is turned on to the end of the current oscillation decay.

[0008] The waveform partitioning module divides the complete waveform into five characteristic time zones based on the abrupt change point of the first-order rate of change in the complete transient time series data. The five characteristic time zones are, in order, the potential plateau region, the current jump region, the voltage collapse region, the current tail region, and the residual oscillation region, and outputs the boundary identifier of the characteristic time zone for each switching cycle.

[0009] The feature extraction module extracts the local slope variation, second derivative zero-crossing density, and short-time energy entropy of the waveform in each feature time zone. It combines the local slope variation, second derivative zero-crossing density, and short-time energy entropy extracted from all feature time zones in the same switching cycle into a high-dimensional morphological feature group and outputs the high-dimensional morphological feature group of the current switching cycle.

[0010] The blind source decomposition module performs non-negative matrix blind source decomposition on the high-dimensional morphological feature group obtained from multiple consecutive switching cycles. During the decomposition process, mutual information is minimized to force each decomposition source to be independent of each other, and three independent source components corresponding to threshold drift, conduction increment and thermal resistance change are separated from it, and the time series of the three independent source components are output.

[0011] The trend analysis and alarm module performs sliding median filtering and exponential weighted trend locking on the time series of the three independent source components to obtain the relative drift of the threshold voltage, the relative increment of the on-resistance, and the relative change of the junction thermal impedance. It then compares each quantity with its preset maintenance threshold and outputs the corresponding maintenance prompt signal based on the comparison result.

[0012] As a further aspect of the present invention: the characteristic time zone boundary identifier for each switching cycle output specifically includes:

[0013] The first-order rate of change of the collected gate potential timing data is calculated point by point to obtain the first-order rate of change sequence. Then, the first-order rate of change sequence is locally accumulated and summed to obtain the cumulative energy curve.

[0014] The inflection point where the slope of the cumulative energy curve changes from steep to gentle is detected, and four boundary points are output in chronological order: the end of the potential plateau region, the end of the current jump region, the end of the voltage collapse region, and the end of the current tail region.

[0015] Starting from the last boundary point, search for the time position in the first-order rate of change sequence where the amplitude decays to below the steady-state noise threshold, which serves as the end boundary of the residual oscillation region, and output five characteristic time zone boundary identifiers for each switching cycle.

[0016] As a further aspect of the present invention: the output process of the high-dimensional morphological feature group is as follows:

[0017] Based on the boundary marker of each characteristic time zone, the slope difference between adjacent points of the gate potential waveform within the characteristic time zone is calculated point by point. The absolute values ​​of all differences are summed and then divided by the time zone length to obtain the local slope variation.

[0018] After applying sliding median filtering to the terminal voltage waveforms within the same time zone, the total number of points where the second derivative changes sign is counted, and then divided by the time width of the time zone, the density of zero-crossing points of the second derivative is obtained.

[0019] The main circuit current waveform within the characteristic time zone is divided into four sub-intervals. The proportion of energy in each sub-interval to the total energy of the entire time zone is calculated. The negative logarithm of each proportion is then summed to obtain the short-time energy entropy.

[0020] The local slope variation, second derivative zero-crossing density, and short-time energy entropy calculated from all characteristic time zones within the same switching cycle are spliced ​​together in a fixed order to output a high-dimensional morphological feature set.

[0021] As a further aspect of the present invention: the sliding median filtering of the terminal voltage waveforms within the same time zone specifically includes:

[0022] Calculate the absolute value sequence of the first-order rate of change of the terminal voltage waveform, and store the absolute value sequence into a fixed-length queue in the order of acquisition.

[0023] Whenever the fixed-length queue is full, the median of all values ​​in the queue is taken as the noise intensity estimate at the current moment. The width of the filtering window is linearly set according to the noise intensity estimate; the larger the noise intensity estimate, the larger the window width.

[0024] After performing sliding median filtering on the primary voltage waveform with a set window width, the fixed-length queue is moved backward by one acquisition step and the earliest value is cleared.

[0025] The entire time zone is processed, and the filtered terminal voltage waveform is output.

[0026] As a further aspect of the present invention: the output of the time sequence of three independent source components specifically includes:

[0027] The high-dimensional morphological feature groups of multiple consecutive switching cycles are arranged into a two-dimensional array in chronological order. The two-dimensional array is used as the input matrix, and a decomposition basis matrix with three sources and a corresponding coefficient matrix are randomly initialized.

[0028] The basis matrix and coefficient matrix are simultaneously decomposed according to the multiplication update rule. After each iteration, the mutual information estimate between each decomposition source is calculated, and the mutual information estimate is added as a penalty term to the update step size of the next iteration, so that the mutual information gradually decreases with each iteration.

[0029] The iteration stops when the decrease in mutual information estimate is less than 1 / 10 of the previous decrease for three consecutive iterations. Each row of the coefficient matrix is ​​then output as the time series sequence of the corresponding independent source components in chronological order.

[0030] As a further aspect of the present invention: the calculation of the mutual information estimate between each decomposition source specifically includes:

[0031] For the coefficient matrix output in this iteration, take any two rows to form a row pair, calculate the joint probability distribution and the marginal probability distribution of the two row vectors in each row pair, multiply the joint probability distribution by the base-2 logarithm of the ratio of the product of the joint probability distribution and the marginal probability distribution, and sum the sum of all row pairs as the original mutual information value.

[0032] Divide the original mutual information value by the total number of iterations completed to obtain the decay coefficient, and then multiply the decay coefficient by a preset mutual information target threshold to obtain the dynamic penalty amount.

[0033] The dynamic penalty is multiplied by the reciprocal of the square root into the update factor of the multiplication update rule for the next iteration, so that the mutual information is forced to approach zero as the number of iterations increases.

[0034] As a further aspect of the present invention: the corresponding operation and maintenance prompt signal output specifically includes:

[0035] For the time series of each independent source component, calculate the absolute difference between the current time and the previous time, and divide the absolute difference by the median of the absolute differences of the last five time points to obtain the local fluctuation ratio.

[0036] The window width of the sliding median filter is linearly set based on the local fluctuation ratio. The larger the local fluctuation ratio, the smaller the window width. Then, the sliding median filter is applied to the sequence based on the window width to output a denoised sequence.

[0037] For the current value in the denoised sequence, calculate its weighted sum with the trend lock value of the previous switching cycle, where the weighting coefficient is equal to the normalized value of the inverse of the local fluctuation ratio, and obtain the trend lock value of the current switching cycle.

[0038] The trend lock value of the current switching cycle is used as the corresponding threshold voltage relative drift, on-resistance relative increment, or junction-shell thermal impedance relative change, and is compared with their respective preset maintenance thresholds in turn to output maintenance prompt signals.

[0039] As a further aspect of the present invention: the output denoising sequence specifically includes:

[0040] A minimum window width and a maximum window width are preset, and the local fluctuation ratio is linearly mapped from the range of zero to the maximum value to the range of the minimum window width to the maximum window width to obtain the window width at the current time.

[0041] Centered on the current moment, take the sequence values ​​corresponding to half the width of the window forward and backward to form a temporary subsequence;

[0042] Sort all values ​​in the temporary subsequence by size, and take the value in the middle position after sorting as the filter output value at the current time.

[0043] Move one sampling step along the time direction and repeat the process to the preset filtered output value until the entire time sequence of independent source components is traversed. Arrange all filtered output values ​​in time order and output the denoised sequence.

[0044] The beneficial effects of this invention are:

[0045] (1) By dividing the switching transient waveform into five characteristic time zones and extracting the local slope rate, the zero-crossing density of the second derivative and the short-time energy entropy, and then combining the non-negative matrix blind source decomposition with mutual information minimization, it is possible to separate the three mutually coupled degradation parameters of threshold drift, conduction increment and thermal resistance change from a single waveform without relying on an accurate physical model. This effectively avoids the error crosstalk caused by strong parameter correlation in the middle stage of power unit aging caused by the traditional linear decoupling method, so that the decoupling result maintains stable physical orientation in different aging stages.

[0046] (2) By introducing sliding median filtering based on the local fluctuation ratio and exponential weighted trend locking, the filtering window width is adaptively adjusted in the opposite direction according to the intensity of signal fluctuation, and the weighting coefficient in trend locking changes monotonically with the local fluctuation ratio, so that the degradation parameter sequence is sufficiently smoothed in the stable stage and keeps fast tracking in the abrupt stage. Thus, without increasing the sampling hardware cost, the robustness of the operation and maintenance prompt signal under random operating conditions is improved, and false triggering caused by noise or load fluctuation is reduced. Attached Figure Description

[0047] The invention will now be further described with reference to the accompanying drawings.

[0048] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0049] 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.

[0050] Please see Figure 1 As shown, this invention is a power unit dynamic parameter optimization and operation and maintenance management system for a power system, comprising:

[0051] The transient data acquisition module collects complete transient timing data of the gate potential, main circuit current and terminal voltage of the power unit during the switching transition. The complete transient timing data covers the entire time period from before the gate drive is turned on to the end of the current oscillation decay.

[0052] The waveform partitioning module divides the complete waveform into five characteristic time zones based on the abrupt change point of the first-order rate of change in the complete transient time series data. The five characteristic time zones are, in order, the potential plateau region, the current jump region, the voltage collapse region, the current tail region, and the residual oscillation region, and outputs the boundary identifier of the characteristic time zone for each switching cycle.

[0053] The feature extraction module extracts the local slope variation, second derivative zero-crossing density, and short-time energy entropy of the waveform in each feature time zone. It combines the local slope variation, second derivative zero-crossing density, and short-time energy entropy extracted from all feature time zones in the same switching cycle into a high-dimensional morphological feature group and outputs the high-dimensional morphological feature group of the current switching cycle.

[0054] The blind source decomposition module performs non-negative matrix blind source decomposition on the high-dimensional morphological feature group obtained from multiple consecutive switching cycles. During the decomposition process, mutual information is minimized to force each decomposition source to be independent of each other, and three independent source components corresponding to threshold drift, conduction increment and thermal resistance change are separated from it, and the time series of the three independent source components are output.

[0055] The trend analysis and alarm module performs sliding median filtering and exponential weighted trend locking on the time series of the three independent source components to obtain the relative drift of the threshold voltage, the relative increment of the on-resistance, and the relative change of the junction thermal impedance. It then compares each quantity with its preset maintenance threshold and outputs the corresponding maintenance prompt signal based on the comparison result.

[0056] The transient data acquisition module collects complete transient timing data of the gate potential, main circuit current, and terminal voltage of the power unit during the switching transition. The complete transient timing data covers the entire time period from before the gate drive is turned on to the end of the current oscillation decay, specifically including:

[0057] In this embodiment, the power unit uses a silicon carbide metal-oxide-semiconductor field-effect transistor as a switching element.

[0058] To acquire the gate potential, a passive high-impedance differential probe with a bandwidth of 200 MHz and an input capacitance of no more than 2 picofarads is connected in parallel between the gate and the source in the power unit drive circuit. The output of the probe is connected to a four-channel synchronous high-speed data acquisition card with a sampling rate of 100 megahertz per second and a resolution of 12 bits.

[0059] To acquire the main circuit current, a coaxial shunt is connected in series with the drain of the power unit. The shunt has a resistance of 0.01 ohms and a bandwidth of 150 MHz. The voltage signal at its two ends is conditioned by a differential amplifier and then connected to the second channel of the acquisition card.

[0060] To acquire the voltage at the acquisition end, a high-voltage differential probe with a bandwidth of 250 MHz and a maximum input voltage of 1200 V is directly connected between the drain and source terminals, and its output terminal is connected to the third channel of the acquisition card.

[0061] The sampling clocks for all three channels are driven by the same external clock source, ensuring a time synchronization error of less than 1 nanosecond. The trigger signal for acquisition is taken from the leading edge of the drive pulse. The starting point of the acquisition window is set 2 microseconds before the rising edge of the drive pulse to record the static noise baseline before the switch is turned on. The ending point of the acquisition window is set 10 microseconds after the falling edge of the drive pulse. This duration is sufficient to cover the high-frequency oscillation attenuation to the steady-state noise level after the drain current is completely turned off.

[0062] The acquisition card converts the analog signals from the three channels into digital timing data, obtaining 4000 sampling points per switching cycle (100 megaseconds multiplied by 40 microseconds). Among these, 4000 points each are for gate potential data, main circuit current data, and terminal voltage data, which together constitute a complete transient timing data record. All records are stored in local cache in chronological order.

[0063] In the waveform partitioning module, based on the abrupt change points of the first-order rate of change in the complete transient time series data, the complete waveform is divided into five characteristic time zones. These five characteristic time zones are, in order, the potential plateau region, the current jump region, the voltage sag region, the current tail region, and the residual oscillation region. The boundary markers of the characteristic time zones for each switching cycle are output, specifically including:

[0064] The first-order rate of change is calculated point-by-point for the acquired gate potential timing data. Specifically, the gate potential value at the current sampling point is subtracted from the gate potential value at the previous sampling point, and the difference is divided by the time interval between two adjacent sampling points (i.e., the sampling period, which is 10 nanoseconds in this embodiment) to obtain the first-order rate of change at that point. This calculation is performed on all sampling points to obtain a sequence of first-order rates of change. Subsequently, the sequence of first-order rates of change is locally accumulated and summed to obtain the cumulative energy curve. The specific method for local accumulation and summation is as follows: starting from the first sampling point, the absolute values ​​of the first-order rates of change at the current point and all previous points are accumulated point-by-point. The accumulated value corresponding to each sampling point constitutes a point on the cumulative energy curve.

[0065] The inflection point where the slope of the cumulative energy curve gradually decreases is detected. The specific detection method is as follows: Calculate the first derivative of the cumulative energy curve (i.e., the difference in cumulative energy values ​​between two adjacent points divided by the time interval). Set a continuous rise threshold. When the value of the first derivative at three consecutive sampling points is less than 50% of the first derivative value at the previous sampling point, the starting point of that continuous segment is determined as an inflection point. Four inflection points are detected sequentially in chronological order and marked as the end of the potential plateau region, the end of the current jump region, the end of the voltage collapse region, and the end of the current tailing region, respectively. The time position corresponding to each inflection point is the end boundary of that characteristic time zone.

[0066] Starting from the time position corresponding to the fourth inflection point (i.e., the end of the current tail region), the amplitude of the first-order rate of change sequence is searched point by point. A steady-state noise threshold is pre-calculated: the average absolute value of the first-order rate of change of all sampling points within 2 microseconds before the gate drive is turned on (i.e., 2 microseconds before the trigger signal's leading edge to the trigger time) is multiplied by 3 to obtain the steady-state noise threshold. Starting from the beginning, when the absolute value of the first-order rate of change of 10 consecutive sampling points is lower than this steady-state noise threshold, the time position corresponding to the first sampling point that meets the condition is determined as the end boundary of the residual oscillation region. Thus, the starting and ending boundaries of the potential plateau region, current surge region, voltage collapse region, current tail region, and residual oscillation region are completely determined: the starting point of the potential plateau region is the beginning of the acquisition window (2 microseconds before the rising edge of the driving pulse), and the ending point is the first inflection point; the starting point of the current surge region is the first inflection point, and the ending point is the second inflection point; the starting point of the voltage collapse region is the second inflection point, and the ending point is the third inflection point; the starting point of the current tail region is the third inflection point, and the ending point is the fourth inflection point; the starting point of the residual oscillation region is the fourth inflection point, and the ending point is the searched end boundary. The time zone boundary identifiers of the above five characteristics for each switching cycle are output.

[0067] In the feature extraction module, the local slope variation, second derivative zero-crossing density, and short-time energy entropy of the waveform are extracted in each feature time zone. The local slope variation, second derivative zero-crossing density, and short-time energy entropy extracted from all feature time zones within the same switching cycle are combined into a high-dimensional morphological feature set, and the high-dimensional morphological feature set of the current switching cycle is output, specifically including:

[0068] The process of extracting the local slope variation is as follows: Based on the boundary markers of each characteristic time zone, the gate potential waveform data within that time zone is extracted. The gate potential waveform consists of a series of discrete sampling points, with a time interval of 10 nanoseconds between adjacent points. For each pair of adjacent sampling points within the time zone, the gate potential difference between the later and earlier points is calculated. This difference is then divided by 10 nanoseconds to obtain the slope between the two points. Subsequently, the absolute value of the difference between two adjacent slopes is calculated. The absolute values ​​of all adjacent slope differences within the entire time zone are summed to obtain the total variation. Finally, this total variation is divided by the time length of the characteristic time zone (i.e., the difference between the end time and the start time, in nanoseconds), and the quotient is the local slope variation of that characteristic time zone. The local slope variation reflects the overall severity of the curvature of the gate potential waveform within that time zone.

[0069] The sliding median filtering process for the terminal voltage waveform: To suppress high-frequency noise in the terminal voltage waveform, a sliding median filter is first applied to the waveform. The specific steps are as follows:

[0070] The first step is to calculate the absolute value sequence of the first-order rate of change of the terminal voltage waveform. For each sampling point, subtract the terminal voltage value from the terminal voltage value of the previous sampling point, take the absolute value, and then divide it by 10 nanoseconds to obtain the absolute value of the first-order rate of change at that point. After calculating all sampling points, a one-dimensional absolute value sequence is obtained.

[0071] The second step is to set up a fixed-length queue with a fixed length of 51 values. Values ​​from the absolute value sequence are sequentially stored into this queue according to the collection order. Each time a value is stored, the queue moves forward one position. When the queue is full (51 values), proceed to the next step.

[0072] The third step is to take the median of all 51 values ​​in the current queue as the noise intensity estimate for the current moment. The median is obtained by arranging the 51 values ​​in the queue in ascending order and taking the 26th value (the value at the middle position) as the median.

[0073] The fourth step is to linearly set the width of the filtering window based on the estimated noise intensity value. Let the estimated noise intensity value be... The unit is volts per nanosecond. A minimum window width is preset. Three sampling points, one maximum window width There are 21 sampling points, and a reference noise upper limit is set. The value is 0.5 volts per nanosecond. The formula for calculating the filter window width W is as follows:

[0074] ;

[0075] in, Take 3. Take 21, Take 0.5. When When equal to 0, Equals 3; when When greater than or equal to 0.5, Equals 21; when When W is between 0 and 0.5, it takes values ​​between 3 and 21 in a linear proportion. The calculated... Round to the nearest integer, and if If the number is even, add 1 to make it odd, to ensure that the window has a unique center point.

[0076] Step 5: Use the calculated window width Perform sliding median filtering on the original terminal voltage waveform: taking samples forward and backward from the current sampling point as the center. -1) / 2 sampling points form a temporary subsequence. All values ​​in the subsequence are sorted by size, and the value in the middle position after sorting is taken as the filtered output value of the current point.

[0077] The sixth step involves moving the fixed-length queue forward by one sampling step, discarding the earliest value stored in the queue, and adding the next new value to the end of the queue. Steps three through five are then repeated until all sampling points within the entire characteristic time zone have been traversed. The final output is the filtered terminal voltage waveform.

[0078] Extraction of the zero-crossing density of the second derivative: For the filtered terminal voltage waveform described above, calculate its second derivative point by point. The calculation method for the second derivative is as follows: first calculate the first derivative of two adjacent points, then calculate the difference between the two adjacent first derivatives and divide by the time interval. For each sampling point, determine whether the sign of its second derivative is opposite to that of the second derivative of the previous sampling point. If opposite, mark the point as a sign-changing point. Count the total number of sign-changing points in the entire characteristic time zone. Then divide the total number by the time width of the characteristic time zone (in microseconds), and the quotient is the zero-crossing density of the second derivative in the characteristic time zone. This density value describes the frequency of changes in the convexity of the terminal voltage waveform.

[0079] Short-time energy entropy extraction: The main circuit current waveform within the characteristic time zone is divided into four equal sub-intervals. Each sub-interval has an equal time length, equal to one-quarter of the total time length of the characteristic time zone. For each sub-interval, the sum of the squares of the current values ​​at all sampling points within that sub-interval is calculated as the energy of that sub-interval. The energies of the four sub-intervals are then added together to obtain the total energy of the entire characteristic time zone. Subsequently, the proportion of energy in each sub-interval to the total energy is calculated, denoted as the first sub-interval. The proportion of each sub-interval is ( =1,2,3,4). Then, according to the definition of information entropy, the short-time energy entropy H is calculated, and the calculation formula is as follows: ;in, Let be the ratio of the energy of each subinterval to the total energy, and each... All are greater than 0; when When equal to 0, it is stipulated that During the calculation, first calculate the values ​​separately. , , , Then calculate each one separately. Multiply by base 2 The logarithm of the product of the four products is taken as the negative value, which gives the short-time energy entropy. This entropy value reflects the uniformity of the energy distribution of the current waveform along the time axis.

[0080] The combined output of the high-dimensional morphological feature set: For the five characteristic time zones (potential plateau region, current rise region, voltage collapse region, current tail region, and residual oscillation region) within a switching cycle, a local slope variation, a second derivative zero-crossing density, and a short-time energy entropy are extracted for each characteristic time zone using the method described above. Therefore, a total of 15 values ​​are obtained for each switching cycle (5 time zones × 3 feature values). These values ​​are concatenated in a fixed order: first the local slope variation, second derivative zero-crossing density, and short-time energy entropy of the potential plateau region; then the three feature values ​​of the current rise region; and so on, finally the three feature values ​​of the residual oscillation region. These 15 values ​​together constitute a high-dimensional morphological feature set, which serves as the output of the current switching cycle.

[0081] In the blind source decomposition module, non-negative matrix blind source decomposition is performed on the high-dimensional morphological feature groups obtained from multiple consecutive switching cycles. During the decomposition process, mutual information minimization forces each decomposition source to be independent, separating three independent source components corresponding to threshold drift, conduction increment, and thermal resistance change. The time series sequences of the three independent source components are output, specifically including:

[0082] Construction of the input and initial matrices: The high-dimensional morphological feature sets obtained from 300 consecutive switching cycles are arranged chronologically. Each switching cycle's high-dimensional morphological feature set contains 15 values. These values ​​are arranged in rows, with each row corresponding to the 15 feature values ​​of one switching cycle, resulting in 300 rows and forming a 300-row, 15-column two-dimensional array, which serves as the input matrix. The number of independent sources to be separated is set to 3. A 300-row, 3-column coefficient matrix and a 3-row, 15-column decomposition basis matrix are randomly generated, where each element is a random decimal between 0 and 1, and all elements are not less than 0, satisfying the non-negativity constraint.

[0083] The iterative process under the multiplicative update rule: Following the multiplicative update rule for nonnegative matrix factorization, both the decomposition basis matrix and the coefficient matrix are iteratively updated simultaneously. Each iteration involves: first, reconstructing an approximate input matrix based on the product of the current decomposition basis matrix and the coefficient matrix; calculating the difference between the input matrix and the reconstructed matrix; and then adjusting each element in both the basis matrix and the coefficient matrix according to the difference value by multiplying the original element by an update factor. This update factor is obtained by appropriately transforming the ratio of the input matrix to the reconstructed matrix, ensuring that the adjusted elements remain nonnegative. Each iteration completes the update of all elements.

[0084] Calculation of mutual information estimates: After each iteration, calculate the mutual information estimates between each decomposition source for the currently updated coefficient matrix (300 rows and 3 columns). The specific calculation process is as follows: Take any two columns in the coefficient matrix (e.g., the first and second columns, the first and third columns, the second and third columns). For each pair of columns, consider the two values ​​in the same row of the two columns as a pair. Classify and statistically analyze all different value combinations of this pair appearing in the 300 rows, and divide the frequency of each combination by 300 to obtain the joint probability distribution. At the same time, separately analyze the distribution of values ​​for each column, and divide the frequency of each value appearing by 300 to obtain the marginal probability distribution. For each value combination, calculate the logarithm of the joint probability distribution multiplied by base 2 (the joint probability distribution divided by the product of the two marginal probabilities distributions). Sum the above calculation results for all value combinations to obtain the mutual information value of that column pair. Add the mutual information values ​​of the three column pairs to obtain the original mutual information value.

[0085] Generation and application of dynamic penalty: Let the total number of iterations completed be T (T=1 after the first iteration, T=2 after the second iteration, and so on). Divide the original mutual information value by T to obtain the decay coefficient. Preset a target threshold for mutual information, which is set to 0.001. Multiply the decay coefficient by the target threshold to obtain the dynamic penalty. Calculate the square root of the dynamic penalty and then take its reciprocal to obtain the penalty factor. In the multiplicative update rule of the next iteration, multiply the update factor by the penalty factor. Since the penalty factor gradually adjusts with the number of iterations, the mutual information is forced to approach zero during the iteration process. Specifically, when the original mutual information is large, the dynamic penalty is large, and its square root reciprocal is small, thereby reducing the update factor and suppressing mutual information; when the mutual information gradually decreases, the dynamic penalty becomes smaller, the penalty factor approaches 1, and the iteration resumes the normal update speed.

[0086] Iteration Stopping Condition and Independent Source Component Output: After each iteration, the mutual information between each column pair in the current coefficient matrix is ​​calculated, and the decrease in mutual information relative to the previous iteration is recorded (i.e., the difference between the previous mutual information and the current mutual information divided by the previous mutual information). Iteration stops when the decrease in mutual information is less than 1 / 10 of the previous decrease in three consecutive iterations. At this point, the first column of the coefficient matrix is ​​extracted in row order (i.e., in time order of the switching cycle) to form the time series sequence of the first independent source component; the second column is extracted to form the time series sequence of the second independent source component; and the third column is extracted to form the time series sequence of the third independent source component. The length of each of the three time series is 300, corresponding to the relative intensity of threshold drift, conduction increment, and thermal resistance change, respectively.

[0087] In the trend analysis and alarm module, the time series of the three independent source components are subjected to sliding median filtering and exponential weighted trend locking to obtain the relative drift of the threshold voltage, the relative increment of the on-resistance, and the relative change of the junction-shell thermal impedance. Each quantity is compared with its preset maintenance threshold, and the corresponding maintenance prompt signal is output based on the comparison result, specifically including:

[0088] Calculation of the local fluctuation ratio: For any time series sequence of an independent source component (values ​​corresponding to 300 switching cycles), starting from the second value, calculate the absolute difference between the current value and the previous value. Simultaneously, taking the absolute difference of the five most recent values ​​backward from the current value (if there are fewer than five, use the actual number), sort these five absolute differences in ascending order, and take the third value in the middle of the sorted sequence as the median. Then divide the current absolute difference by this median; the quotient is the local fluctuation ratio at the current time. If the median is zero, the local fluctuation ratio is directly set to a preset large number (e.g., 100). This ratio reflects the severity of local fluctuations in the sequence at the current time.

[0089] Setting the sliding median filter window width: A minimum window width of 3 sampling points and a maximum window width of 21 sampling points are preset. The possible range of local fluctuation ratio is defined as 0 to 10, with a value greater than 10 calculated as 10. A linear mapping method is used: the local fluctuation ratio range of 0 to 10 is linearly mapped to the range of minimum window width 3 to maximum window width 21. The specific mapping rule is as follows: when the local fluctuation ratio is 0, the window width is 21; when the local fluctuation ratio is 10, the window width is 3; for values ​​between 0 and 10, the window width is equal to 21 minus (the local fluctuation ratio multiplied by 1.8), then rounded to the nearest integer, ensuring that the final window width is odd (add 1 if it is even). This yields the filter window width at the current moment. This rule ensures that the larger the local fluctuation ratio (i.e., the more drastic the signal change), the smaller the window width to preserve details; and the smaller the local fluctuation ratio (i.e., the smoother the signal), the larger the window width to adequately smooth noise.

[0090] Generation of the denoising sequence: Centered on the current time step, take half the window width (rounded down) of sampling points forward and half the window width (rounded down) of sampling points backward, along with the current time step itself, to form a temporary subsequence. For example, with a window width of 11, take 5 forward, 5 backward, and 5 forward, for a total of 11 values. Arrange all values ​​in this temporary subsequence in ascending order, and take the value at the middle position (i.e., the (window width + 1) / 2th value) as the filtered output value for the current time step. Then, move one sampling step along the time direction (i.e., move to the next time step), repeating the process of setting the window width and taking the median, until the entire time sequence of the independent source components is traversed. Arrange the filtered output values ​​of all time steps in chronological order to obtain the denoising sequence.

[0091] Calculation of Exponentially Weighted Trend Lock Value: For the denoised sequence, the trend lock value for the current switching cycle is calculated starting from the second value. Let the trend lock value of the previous switching cycle (the trend lock value of the first moment is directly taken as the first value of the denoised sequence), the denoised sequence value at the current moment be the current value, and the local fluctuation ratio at the current moment be denoted as R. First, calculate the weighting coefficient: use the reciprocal of R as the basic weight; if R is zero, directly set the basic weight to 1. Then, normalize the basic weights of all moments, but here we simplify by directly using the normalized value: the weighting coefficient is equal to the reciprocal of R divided by (the reciprocal of R plus 1), so that the weighting coefficient is between 0 and 1. Specifically, the weighting coefficient is equal to 1 divided by (R plus 1). Then, calculate the trend lock value for the current switching cycle: multiply the current value by the weighting coefficient, and add the trend lock value of the previous switching cycle multiplied by (1 minus the weighting coefficient). This calculation method results in a higher weight for the current value and faster trend tracking when the local fluctuation ratio is larger (i.e., the signal changes drastically), and a higher weight for historical trends when the local fluctuation ratio is smaller (the signal is stable), resulting in smoother filtering.

[0092] The maintenance prompt signals are output as follows: The trend lock value of the current switching cycle is used as the relative drift of the threshold voltage, the relative increment of the on-resistance, or the relative change of the junction-case thermal impedance, respectively. For each quantity, a preset maintenance threshold is set: the threshold for the relative drift of the threshold voltage is 0.5 volts, the threshold for the relative increment of the on-resistance is 20% of the initial value, and the threshold for the relative change of the junction-case thermal impedance is 30% of the initial value. The trend lock value of the current switching cycle is compared with the corresponding preset maintenance threshold: if it is greater than or equal to the threshold, a "Maintenance Required" prompt signal is output; if it is less than the threshold but greater than 80% of the threshold, a "Observe Carefully" prompt signal is output; if it is less than 80% of the threshold, a "Normal Operation" prompt signal is output. These three quantities are judged independently, and corresponding maintenance prompt signals are output for maintenance personnel's reference.

[0093] The working principle of this invention is as follows: Complete transient timing data of the gate potential, main circuit current, and terminal voltage of the power unit during the switching transition process is collected from before the gate drive is turned on until the current oscillation decays. Based on the abrupt change points of the first-order rate of change in this data, the complete waveform is sequentially divided into five characteristic time zones: a potential plateau region, a current jump region, a voltage collapse region, a current tail region, and a residual oscillation region. The boundary markers of the characteristic time zones for each switching cycle are output. Within each characteristic time zone, the local slope rate of change, the zero-crossing density of the second derivative, and the short-time energy entropy of the waveform are extracted and combined into a high-dimensional morphological feature group. Then, non-negative matrix blind source decomposition is performed on the high-dimensional morphological feature group for multiple consecutive switching cycles. During the decomposition process, mutual information minimization forces each decomposition source to be independent, separating the timing sequences of three independent source components corresponding to threshold drift, conduction increment, and thermal resistance change. Sliding median filtering and exponential weighted trend locking are applied to the timing sequences of the three independent source components to obtain the relative drift of the threshold voltage, the relative increment of the conduction resistance, and the relative change of the junction-shell thermal resistance. After comparing these with preset maintenance thresholds, the corresponding maintenance prompt signal is output.

[0094] 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 power unit dynamic parameter optimization operation and maintenance management system for a power system, characterized in that, include: The transient data acquisition module collects complete transient timing data of the gate potential, main circuit current and terminal voltage of the power unit during the switching transition. The complete transient timing data covers the entire time period from before the gate drive is turned on to the end of the current oscillation decay. The waveform partitioning module divides the complete waveform into five characteristic time zones based on the abrupt change point of the first-order rate of change in the complete transient time series data. The five characteristic time zones are, in order, the potential plateau region, the current jump region, the voltage collapse region, the current tail region, and the residual oscillation region, and outputs the boundary identifier of the characteristic time zone for each switching cycle. The feature extraction module extracts the local slope variation, second derivative zero-crossing density, and short-time energy entropy of the waveform in each feature time zone. It combines the local slope variation, second derivative zero-crossing density, and short-time energy entropy extracted from all feature time zones in the same switching cycle into a high-dimensional morphological feature group and outputs the high-dimensional morphological feature group of the current switching cycle. The blind source decomposition module performs non-negative matrix blind source decomposition on the high-dimensional morphological feature group obtained from multiple consecutive switching cycles. During the decomposition process, mutual information is minimized to force each decomposition source to be independent of each other, and three independent source components corresponding to threshold drift, conduction increment and thermal resistance change are separated from it, and the time series of the three independent source components are output. The trend analysis and alarm module performs sliding median filtering and exponential weighted trend locking on the time series of the three independent source components to obtain the relative drift of the threshold voltage, the relative increment of the on-resistance, and the relative change of the junction thermal impedance. It then compares each quantity with its preset maintenance threshold and outputs the corresponding maintenance prompt signal based on the comparison result.

2. The power unit dynamic parameter optimization operation and maintenance management system for a power system according to claim 1, characterized in that, The characteristic time zone boundary identifier for each switching cycle output specifically includes: The first-order rate of change of the collected gate potential timing data is calculated point by point to obtain the first-order rate of change sequence. Then, the first-order rate of change sequence is locally accumulated and summed to obtain the cumulative energy curve. The inflection point where the slope of the cumulative energy curve changes from steep to gentle is detected, and four boundary points are output in chronological order: the end of the potential plateau region, the end of the current jump region, the end of the voltage collapse region, and the end of the current tail region. Starting from the last boundary point, search for the time position in the first-order rate of change sequence where the amplitude decays to below the steady-state noise threshold, which serves as the end boundary of the residual oscillation region, and output five characteristic time zone boundary identifiers for each switching cycle.

3. The power unit dynamic parameter optimization operation and maintenance management system for a power system according to claim 1, characterized in that, The output process of the high-dimensional morphological feature group is as follows: Based on the boundary marker of each characteristic time zone, the slope difference between adjacent points of the gate potential waveform within the characteristic time zone is calculated point by point. The absolute values ​​of all differences are summed and then divided by the time zone length to obtain the local slope variation. After applying sliding median filtering to the terminal voltage waveforms within the same time zone, the total number of points where the second derivative changes sign is counted, and then divided by the time width of the time zone, the density of zero-crossing points of the second derivative is obtained. The main circuit current waveform within the characteristic time zone is divided into four sub-intervals. The proportion of energy in each sub-interval to the total energy of the entire time zone is calculated. The negative logarithm of each proportion is then summed to obtain the short-time energy entropy. The local slope variation, second derivative zero-crossing density, and short-time energy entropy calculated from all characteristic time zones within the same switching cycle are spliced ​​together in a fixed order to output a high-dimensional morphological feature set.

4. The power unit dynamic parameter optimization operation and maintenance management system for a power system according to claim 3, characterized in that, The sliding median filtering of the terminal voltage waveforms within the same time zone specifically includes: Calculate the absolute value sequence of the first-order rate of change of the terminal voltage waveform, and store the absolute value sequence into a fixed-length queue in the order of acquisition. Whenever the fixed-length queue is full, the median of all values ​​in the queue is taken as the noise intensity estimate at the current moment. The width of the filtering window is linearly set according to the noise intensity estimate; the larger the noise intensity estimate, the larger the window width. After performing sliding median filtering on the primary voltage waveform with a set window width, the fixed-length queue is moved backward by one acquisition step and the earliest value is cleared. The entire time zone is processed, and the filtered terminal voltage waveform is output.

5. The power unit dynamic parameter optimization operation and maintenance management system for a power system according to claim 1, characterized in that, The output time series of three independent source components specifically includes: The high-dimensional morphological feature groups of multiple consecutive switching cycles are arranged into a two-dimensional array in chronological order. The two-dimensional array is used as the input matrix, and a decomposition basis matrix with three sources and a corresponding coefficient matrix are randomly initialized. The basis matrix and coefficient matrix are simultaneously decomposed according to the multiplication update rule. After each iteration, the mutual information estimate between each decomposition source is calculated, and the mutual information estimate is added as a penalty term to the update step size of the next iteration, so that the mutual information gradually decreases with each iteration. The iteration stops when the decrease in mutual information estimate is less than 1 / 10 of the previous decrease for three consecutive iterations. Each row of the coefficient matrix is ​​then output as the time series sequence of the corresponding independent source components in chronological order.

6. The power unit dynamic parameter optimization operation and maintenance management system for a power system according to claim 5, characterized in that, The calculation of the mutual information estimate between each decomposition source specifically includes: For the coefficient matrix output in this iteration, take any two rows to form a row pair, calculate the joint probability distribution and the marginal probability distribution of the two row vectors in each row pair, multiply the joint probability distribution by the base-2 logarithm of the ratio of the product of the joint probability distribution and the marginal probability distribution, and sum the results of all row pairs as the original mutual information value. Divide the original mutual information value by the total number of iterations completed to obtain the decay coefficient, and then multiply the decay coefficient by a preset mutual information target threshold to obtain the dynamic penalty amount. The dynamic penalty is multiplied by the reciprocal of the square root into the update factor of the multiplication update rule for the next iteration, so that the mutual information is forced to approach zero as the number of iterations increases.

7. The power unit dynamic parameter optimization operation and maintenance management system for a power system according to claim 1, characterized in that, The corresponding maintenance prompt signals for the output specifically include: For the time series of each independent source component, calculate the absolute difference between the current time and the previous time, and divide the absolute difference by the median of the absolute differences of the last five time points to obtain the local fluctuation ratio. The window width of the sliding median filter is linearly set based on the local fluctuation ratio. The larger the local fluctuation ratio, the smaller the window width. Then, the sliding median filter is applied to the sequence based on the window width to output a denoised sequence. For the current value in the denoised sequence, calculate its weighted sum with the trend lock value of the previous switching cycle, where the weighting coefficient is equal to the normalized value of the inverse of the local fluctuation ratio, and obtain the trend lock value of the current switching cycle. The trend lock value of the current switching cycle is used as the corresponding threshold voltage relative drift, on-resistance relative increment, or junction-shell thermal impedance relative change, and is compared with their respective preset maintenance thresholds in turn to output maintenance prompt signals.

8. The power unit dynamic parameter optimization operation and maintenance management system for a power system according to claim 7, characterized in that, The output denoising sequence specifically includes: A minimum window width and a maximum window width are preset, and the local fluctuation ratio is linearly mapped from the range of zero to the maximum value to the range of the minimum window width to the maximum window width to obtain the window width at the current time. Centered on the current moment, take the sequence values ​​corresponding to half the width of the window forward and backward to form a temporary subsequence; Sort all values ​​in the temporary subsequence by size, and take the value in the middle position after sorting as the filter output value at the current time. Move one sampling step along the time direction and repeat the process to the preset filtered output value until the entire time sequence of independent source components is traversed. Arrange all filtered output values ​​in time order and output the denoised sequence.

Images