Optimization method of double pulse test and training method of test parameter generation model

Abstract

Embodiments of the present application provide a kind of double pulse test optimization method and the training method of test parameter generation model.The method comprises: obtaining the working condition parameter and index parameter of inverter;Wherein, working condition parameter and index parameter are the corresponding relationship of the paired data obtained under the same double pulse test condition;Characteristic extraction processing is carried out to working condition parameter and index parameter, and feature vector is obtained;The feature vector is input into initial intelligent model, the feature vector is processed based on initial intelligent model, and the parameters of initial intelligent model are optimized by weighting objective function, and test parameter generation model is obtained;Wherein, test parameter generation model is used to process the working condition parameter and index parameter of inverter and obtain test parameter;Test parameter is used to complete double pulse test of inverter.The method is used to reach the effect of improving the accuracy and execution efficiency of double pulse test.

Description

Technical Field

[0001] This application relates to the field of electronic device testing technology, and in particular to an optimization method for double-pulse testing and a training method for a test parameter generation model. Background Technology

[0002] The double-pulse test simulates the actual switching process by applying two sets of pulse signals to the inverter, thereby acquiring the switching waveform and analyzing key parameters.

[0003] In the existing technology, testers set initial test parameters based on experience; then they start the test, acquire the switching waveform using an oscilloscope, and analyze dynamic response indicators such as voltage overshoot, oscillation amplitude, and switching losses; if the test results do not meet the requirements, one or more parameters need to be manually adjusted, and the iterative test process is repeated until acceptable waveform quality and loss data are obtained.

[0004] However, the above method relies on experience for parameter adjustment, and the test results of different testers or at different time points under the same working conditions can vary significantly, resulting in low testing efficiency. Summary of the Invention

[0005] This application provides an optimization method for bipulse testing and a training method for a test parameter generation model, in order to improve the accuracy and execution efficiency of bipulse testing.

[0006] In a first aspect, embodiments of this application provide a training method for a test parameter generation model, comprising: acquiring the operating condition parameters and index parameters of an inverter; wherein the operating condition parameters and index parameters are paired data with a one-to-one correspondence acquired under the same double-pulse test conditions; performing feature extraction processing on the operating condition parameters and index parameters to obtain feature vectors; inputting the feature vectors into an initial intelligent model, processing the feature vectors based on the initial intelligent model, and optimizing the parameters of the initial intelligent model through a weighted objective function to obtain a test parameter generation model; wherein the test parameter generation model is used to process the operating condition parameters and index parameters of the inverter to obtain test parameters; and the test parameters are used to complete the double-pulse test of the inverter.

[0007] In one possible implementation, feature extraction processing is performed on the operating condition parameters and index parameters to obtain feature vectors, including: performing feature extraction processing on the operating condition parameters to obtain an operating condition feature vector; performing instantaneous feature extraction processing on the index parameters to obtain an instantaneous feature vector; wherein, the operating condition feature vector characterizes the static operating state and working conditions of the inverter under a double-pulse test; the instantaneous feature vector characterizes the transient change characteristics of the electrical parameters of the inverter during the double-pulse test; the operating condition feature vector and the instantaneous feature vector are fused to obtain a fused feature vector; and the fused feature vector is normalized to obtain a feature vector.

[0008] In one possible implementation, feature extraction processing is performed on the operating parameters to obtain an operating feature vector, including: performing time-domain statistical feature extraction processing on the operating parameters to obtain statistical features; performing time-series analysis processing on the operating parameters to extract their dynamic change trend to obtain dynamic features; wherein, the statistical features characterize at least one of the mean, variance, and peak value of the operating parameters; the dynamic features characterize the slope or frequency component of the operating parameters as they evolve over time; and the statistical features and dynamic features are fused to determine the operating feature vector.

[0009] In one possible implementation, the working condition feature vector and the instantaneous feature vector are fused to obtain a fused feature vector, including: using an attention mechanism to determine a first weight for each feature element in the working condition feature vector and a second weight for each feature element in the instantaneous feature vector; weighting the working condition feature vector according to the first weight to obtain a weighted working condition feature vector, and weighting the instantaneous feature vector according to the second weight to obtain a weighted instantaneous feature vector; and concatenating the weighted working condition feature vector and the weighted instantaneous feature vector to generate a fused feature vector.

[0010] In one possible implementation, the feature vector is associated with a corresponding test parameter label, which is a test parameter reference value that corresponds one-to-one with the operating condition parameter and the index parameter. The feature vector is input into an initial intelligent model, processed based on the initial intelligent model, and the parameters of the initial intelligent model are optimized using a weighted objective function to obtain a test parameter generation model. This includes: inputting the feature vector into the initial intelligent model; processing the feature vector based on the initial intelligent model to obtain initial test parameters; determining a target value based on the initial test parameters, the test parameter labels of the feature vector, and the weighted objective function; and optimizing the parameters in the initial intelligent model using a chain rule based on the target value to obtain the test parameter generation model.

[0011] In one possible implementation, the feature vector is associated with corresponding test parameter labels, which are test parameter reference values ​​that correspond one-to-one with the operating condition parameters and index parameters. The method further includes: determining the predicted switching loss based on the initial test parameters and the test parameter labels in the feature vector; determining the switching loss error based on the predicted switching loss and the switching loss in the feature vector; determining the index oscillation amplitude based on the index parameters in the feature vector and a preset index threshold; and determining a weighted objective function with the goal of minimizing the switching loss error and the index oscillation amplitude.

[0012] Secondly, embodiments of this application provide an optimized method for double-pulse testing, comprising: acquiring the operating condition parameters of an inverter; performing matching processing on the operating condition parameters in a preset database to determine initial test parameters; performing double-pulse testing on the inverter based on the operating condition parameters and the initial test parameters to obtain initial index parameters; and generating a model based on the operating condition parameters, the initial index parameters, and the test parameters to determine the target index parameters corresponding to the operating condition parameters; wherein the test parameter generation model is trained using the first aspect and / or various possible training methods described above.

[0013] In one possible implementation, the target index parameter corresponding to the operating condition parameter is determined based on the operating condition parameter, initial index parameter, and test parameter generation model. This includes repeatedly executing the following steps until a preset condition is met: inputting the operating condition parameter and the i-th initial index parameter into the test parameter generation model to obtain the i-th test parameter output by the test parameter generation model; performing a double-pulse test on the inverter based on the operating condition parameter and the i-th test parameter to obtain the (i+1)-th initial index parameter; where the first initial index parameter is an initial index parameter; i is a positive integer greater than or equal to 1; and determining the value of i plus 1; wherein the initial index parameter obtained when the preset condition is met is used to determine the target index parameter corresponding to the operating condition parameter.

[0014] Thirdly, embodiments of this application provide a training apparatus for a test parameter generation model, comprising: a first acquisition module for acquiring the operating condition parameters and index parameters of an inverter; wherein the operating condition parameters and index parameters are paired data with a one-to-one correspondence acquired under the same double-pulse test conditions; a first processing module for performing feature extraction processing on the operating condition parameters and index parameters to obtain feature vectors; and a training module for inputting the feature vectors into an initial intelligent model, processing the feature vectors based on the initial intelligent model, and optimizing the parameters of the initial intelligent model through a weighted objective function to obtain a test parameter generation model; wherein the test parameter generation model is used to process the operating condition parameters and index parameters of the inverter to obtain test parameters; and the test parameters are used to complete the double-pulse test of the inverter.

[0015] In one possible implementation, the processing module includes: a processing submodule, used to perform feature extraction processing on the operating condition parameters to obtain an operating condition feature vector; and to perform instantaneous feature extraction processing on the index parameters to obtain an instantaneous feature vector; wherein the operating condition feature vector characterizes the static operating state and working conditions of the inverter under a double-pulse test; and the instantaneous feature vector characterizes the transient change characteristics of the electrical parameters of the inverter during the double-pulse test; and a fusion module, used to perform fusion processing on the operating condition feature vector and the instantaneous feature vector to obtain a fused feature vector; and to perform normalization processing on the fused feature vector to obtain a feature vector.

[0016] In one possible implementation, the processing submodule includes: performing time-domain statistical feature extraction processing on the operating parameters to obtain statistical features; performing time-series analysis processing on the operating parameters to extract their dynamic change trends to obtain dynamic features; wherein, the statistical features characterize at least one of the mean, variance, and peak value of the operating parameters; the dynamic features characterize the slope or frequency component of the operating parameters as they evolve over time; and performing fusion processing on the statistical features and dynamic features to determine the operating feature vector.

[0017] In one possible implementation, the fusion module includes: employing an attention mechanism to determine a first weight for each feature element in the working condition feature vector and a second weight for each feature element in the instantaneous feature vector; weighting the working condition feature vector according to the first weight to obtain a weighted working condition feature vector, and weighting the instantaneous feature vector according to the second weight to obtain a weighted instantaneous feature vector; and concatenating the weighted working condition feature vector and the weighted instantaneous feature vector to generate a fused feature vector.

[0018] In one possible implementation, the feature vector is associated with a corresponding test parameter label, which is a test parameter reference value that corresponds one-to-one with the operating condition parameter and the index parameter; the training module includes: inputting the feature vector into an initial intelligent model, processing the feature vector based on the initial intelligent model to obtain initial test parameters; determining a target value based on the initial test parameters, the test parameter labels of the feature vector, and a weighted objective function; and optimizing the parameters in the initial intelligent model according to the target value using a chain rule to obtain a test parameter generation model.

[0019] In one possible implementation, the feature vector is associated with corresponding test parameter labels, which are test parameter reference values ​​that correspond one-to-one with the operating condition parameters and index parameters. The device further includes: determining the predicted switching loss based on the initial test parameters and the test parameter labels in the feature vector; determining the switching loss error based on the predicted switching loss and the switching loss in the feature vector; determining the index oscillation amplitude based on the index parameters in the feature vector and a preset index threshold; and determining a weighted objective function with the goal of minimizing the switching loss error and the index oscillation amplitude.

[0020] Fourthly, embodiments of this application provide an optimization apparatus for double-pulse testing, comprising: an acquisition module for acquiring the operating condition parameters of an inverter; performing matching processing on the operating condition parameters in a preset database to determine initial test parameters; a processing module for performing double-pulse testing on the inverter based on the operating condition parameters and the initial test parameters to obtain initial index parameters; and a determination module for generating a model based on the operating condition parameters, the initial index parameters, and the test parameters to determine the target index parameters corresponding to the operating condition parameters; the test parameter generation model is trained through the third aspect and / or various possible training devices described above.

[0021] In one possible implementation, the determining module includes: repeatedly executing the following steps until a preset condition is met: inputting the operating condition parameters and the i-th initial index parameter into the test parameter generation model to obtain the i-th test parameter output by the test parameter generation model; performing a double-pulse test on the inverter based on the operating condition parameters and the i-th test parameter to obtain the (i+1)-th initial index parameter; wherein, the first initial index parameter is an initial index parameter; i is a positive integer greater than or equal to 1; and determining the value of i plus 1; wherein, the initial index parameter obtained when the preset condition is met is used to determine the target index parameter corresponding to the operating condition parameters.

[0022] Fifthly, embodiments of this application provide an inverter with an integrated testing module, comprising: an inverter for converting direct current (DC) to alternating current (AC); and a testing module electrically connected to the inverter for executing the training method of the test parameter generation model as described in the first aspect and / or the various possible test parameter generation models of the first aspect, or the optimization method of the double-pulse test as described in the second aspect and / or the various possible double-pulse test models of the second aspect; wherein the testing module comprises: a memory for storing operating condition parameters, index parameters, and test parameter generation models; and a processor for executing the training operation of the test parameter generation model or the optimization operation of the double-pulse test.

[0023] In a sixth aspect, embodiments of this application provide an electronic device, including: a memory and a processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0024] In a seventh aspect, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0025] Eighthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0026] This application provides an optimization method for double-pulse testing and a training method for a test parameter generation model. First, it acquires paired data of inverter operating condition parameters and index parameters with a one-to-one correspondence under the same double-pulse test conditions, ensuring the homogeneity and matching degree of the training data. Then, it performs feature extraction processing on the paired operating condition parameters and index parameters to obtain feature vectors, completing the screening and dimensionality reduction of effective features and eliminating redundant interference information. Subsequently, the feature vectors are input into an initial intelligent model for processing, and the parameters of the initial intelligent model are iteratively optimized through a weighted objective function. Finally, a generation model that can output double-pulse test parameters based on inverter operating conditions and index parameters is obtained. This technical means realizes the intelligent and accurate generation of inverter double-pulse test parameters, significantly reducing the time and labor costs of manual debugging, effectively solving the pain points of traditional test parameters relying on human experience and poor operating condition adaptability, significantly improving the matching degree between test parameters and the actual operating conditions of the inverter, while ensuring the accuracy and execution efficiency of double-pulse testing, and can quickly adapt to the inverter testing needs under different operating conditions. Attached Figure Description

[0027] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0028] Figure 1 A flowchart illustrating a training method for a test parameter generation model provided in this application embodiment. Figure 1 ;

[0029] Figure 2 A flowchart illustrating a training method for a test parameter generation model provided in this application embodiment. Figure 2 ;

[0030] Figure 3 A flowchart illustrating an optimized method for dual-pulse testing provided in an embodiment of this application;

[0031] Figure 4 A schematic diagram of the structure of a training device for a test parameter generation model provided in an embodiment of this application;

[0032] Figure 5 A schematic diagram of the structure of an optimized device for dual-pulse testing provided in an embodiment of this application;

[0033] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0034] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0035] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0036] The double-pulse test is an important experimental method in power electronics for evaluating the dynamic characteristics of inverter switching devices (such as IGBTs and MOSFETs). This test simulates the switching process during actual operation by applying two sets of pulse signals to the inverter, thereby acquiring switching waveforms and analyzing key parameters. This scenario is widely used in new energy power generation systems, electric vehicle drive systems, and industrial motor drive equipment.

[0037] In actual testing, testers need to preset parameters such as pulse width, drive voltage, drive resistance, and pulse interval time based on experience, and monitor the switching waveform in real time using an oscilloscope. Because the switching characteristics of power electronic devices are significantly affected by operating conditions, the test parameters need to be dynamically adjusted to meet the performance requirements under different operating conditions.

[0038] However, current testing procedures rely heavily on manual experience, and parameter adjustments lack unified quantitative guidelines, resulting in low testing efficiency, poor result reproducibility, and difficulty in adapting to the rapid response requirements of complex operating conditions. Furthermore, as power electronic systems evolve towards higher power density and higher efficiency, the optimization requirements for switching losses and dynamic response are becoming increasingly stringent, making traditional manual testing methods insufficient to meet the urgent needs of modern industry for standardized, automated, and intelligent testing.

[0039] Therefore, this application provides an optimized method for dual-pulse testing, which can solve the above-mentioned problems.

[0040] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0041] Figure 1A flowchart illustrating a training method for a test parameter generation model provided in this application embodiment. Figure 1 ,like Figure 1 As shown, the method includes:

[0042] S101. Obtain the inverter's operating condition parameters and index parameters.

[0043] Among them, the operating condition parameters and index parameters are paired data with a one-to-one correspondence obtained under the same double-pulse test conditions.

[0044] For example, an inverter is a power electronic conversion device that converts direct current into alternating current, and it is widely used in fields such as motor drive and new energy grid connection.

[0045] The double-pulse test is a standard method for evaluating the dynamic characteristics of semiconductor devices. It simulates the switching behavior of the power switching transistors (such as IGBTs or SiC MOSFETs) in the inverter under test by applying two consecutive pulse signals.

[0046] Operating parameters refer to the electrical operating conditions that are precisely set or recorded during the double-pulse test, including but not limited to controllable variables such as bus voltage, load current, drive resistance, junction temperature, and pulse width.

[0047] The performance parameters are key results data that characterize the switch performance, which are directly measured or calculated by the testing instrument. These include, but are not limited to, turn-on loss, voltage change rate, current change rate, voltage overshoot, and current spike.

[0048] For example, all completed and fully-fledged double-pulse test records are selected from the historical database of inverter R&D or factory inspection. Each record must contain a unique test identifier.

[0049] Read the host computer log file or experimental record table of the test equipment and extract the corresponding operating condition settings for each record. This includes reading the set bus voltage value from the DC power supply, reading the estimated current case temperature or junction temperature from the temperature control box, reading the set on / off drive resistor value from the pulse generator, and the target load current value corresponding to the pulse width.

[0050] Access the raw waveform data file saved on the oscilloscope. Post-process the voltage and current waveforms using preset mathematical algorithms to calculate key indicators of the switching transients. For example, the energy loss during a single switch is obtained by integrating the product of the instantaneous voltage and current values, and the overshoot voltage is obtained by finding the maximum voltage spike. These calculated values ​​are archived as structured indicator data.

[0051] Using the test identifier as the primary key, the parsed working condition parameter vector and the calculated index parameter vector are strictly joined, and invalid samples with missing or abnormal data on either side are removed. Finally, a structured table is generated, where each row represents an independent training sample, the first few columns are the working condition input features, and the last few columns are the index output labels.

[0052] S102. Perform feature extraction processing on the operating condition parameters and index parameters to obtain feature vectors.

[0053] For example, the operating condition parameters and index parameters are standardized and normalized. The mean and standard deviation of each operating condition parameter and index parameter in the entire training sample set are calculated, and then the Z-score standardization formula is applied to map the data to a distribution with a mean of 0 and a standard deviation of 1. Optionally, for parameters with nonlinear saturation characteristics, logarithmic transformation or Box-Cox transformation can be used to improve the symmetry of the data distribution.

[0054] Interactive and physical meaning features are constructed for the processed operating parameters and index parameters. While the operating parameters are independent variables, their interactions are significant in power semiconductor physics. For example, turn-off losses depend not only on the current magnitude but also strongly on the product of current and voltage, as well as the charging and discharging process of parasitic capacitance. The feature extraction algorithm automatically constructs composite feature terms based on knowledge of power electronics. For the dynamic characteristics of the index parameters, their second-order statistics or rates of change are calculated to capture the instantaneous trends in the switching trajectory.

[0055] Feature selection and dimensionality reduction are performed on the processed operating parameters and index parameters. The Pearson correlation coefficient between each candidate feature and the target output (i.e., the test parameters to be generated) is calculated, and weakly correlated features with absolute correlation values ​​below a threshold are eliminated. Simultaneously, principal component analysis is used to project the high-dimensional feature space onto a low-dimensional orthogonal basis. While retaining more than 95% of variance information, the dozens of features are compressed into a few principal components, ultimately outputting a compact and highly information-dense feature vector. This vector is stored in the data pipeline, ready to be input into the training phase of the generative model.

[0056] S103. Input the feature vector into the initial intelligent model, process the feature vector based on the initial intelligent model, and optimize the parameters of the initial intelligent model through a weighted objective function to obtain the test parameter generation model.

[0057] The test parameter generation model is used to process the inverter's operating parameters and index parameters to obtain test parameters; the test parameters are used to complete the double-pulse test of the inverter.

[0058] For example, feature vectors are input into an initial intelligent model. The initial model performs a series of linear transformations and non-linear activation operations on the feature vectors based on its current internal parameters, namely, the weight matrix and the bias vector, ultimately producing a set of values ​​at the output layer. In the initial stage of training, since the parameters are random values, there is a significant deviation between this set of output values ​​(i.e., the test parameters guessed by the model) and the true test parameter labels.

[0059] The predicted test parameter vector output by the model will be compared with the true test parameter labels in the historical paired data. Here, weighted mean squared error or weighted average absolute percentage error is used as the objective function.

[0060] Optionally, the model reads a pre-defined weight vector. This mechanism forces the model to prioritize the prediction accuracy of high-weight metrics during optimization, even if it means making minor compromises on low-weight metrics.

[0061] Backpropagation and iterative parameter optimization are performed. The scalar loss value calculated from the weighted objective function is used to calculate the gradient vector of the loss with respect to each trainable parameter in the model through an automatic differentiation mechanism. The optimizer algorithm (such as Adam or stochastic gradient descent) fine-tunes and updates the weights and biases of the initial intelligent model according to the gradient direction and the preset learning rate to reduce the loss value in the next iteration. The training model loop will traverse the entire training dataset hundreds or even thousands of times. As the number of iterations increases, the loss value gradually converges to a stable and low value. At this point, the model output test parameters are highly consistent with the historical successful test parameters, and the model state at this time is frozen and exported as the test parameter generation model.

[0062] This application provides a training method for a test parameter generation model. First, it acquires paired data of inverter operating condition parameters and index parameters with a one-to-one correspondence under the same double-pulse test conditions, ensuring the homogeneity and matching degree of the training data. Then, it performs feature extraction processing on the paired operating condition parameters and index parameters to obtain feature vectors, completing the screening and dimensionality reduction of effective features and eliminating redundant interference information. Subsequently, the feature vectors are input into an initial intelligent model for processing, and the parameters of the initial intelligent model are iteratively optimized through a weighted objective function. Finally, a generation model that can output double-pulse test parameters based on inverter operating conditions and index parameters is obtained. This technical means realizes the intelligent and accurate generation of inverter double-pulse test parameters, significantly reducing the time and labor costs of manual debugging, effectively solving the pain points of traditional test parameters relying on human experience and having poor operating condition adaptability. It significantly improves the matching degree between test parameters and the actual operating conditions of the inverter, while ensuring the accuracy and execution efficiency of double-pulse testing, and can quickly adapt to the inverter testing needs under different operating conditions.

[0063] Figure 2A flowchart illustrating a training method for a test parameter generation model provided in this application embodiment. Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the examples, a training method for a test parameter generation model is described in detail, the method including:

[0064] S201. Obtain the inverter's operating condition parameters and index parameters.

[0065] Among them, the operating condition parameters and index parameters are paired data with a one-to-one correspondence obtained under the same double-pulse test conditions.

[0066] For example, this step can be referred to step S101, and will not be described again.

[0067] S202. Perform feature extraction processing on the operating condition parameters to obtain the operating condition feature vector; perform instantaneous feature extraction processing on the index parameters to obtain the instantaneous feature vector.

[0068] Among them, the operating condition feature vector represents the static operating state and working conditions of the inverter under the double-pulse test; the instantaneous feature vector represents the transient change characteristics of the electrical parameters of the inverter during the double-pulse test.

[0069] For example, a list of operating condition parameters is obtained; optionally, the list of operating condition parameters includes, but is not limited to, numerical continuous variables and categorical discrete variables. For continuous variables, Z-score normalization or binning discretization is used to adapt their distribution to the requirements of the model input layer. Subsequently, interaction terms between operating conditions are constructed based on knowledge in the field of power electronics, such as calculating "the square of the bus voltage divided by the gate resistance" to estimate the impact of Miller capacitor discharge rate, or calculating "the normalized increment of junction temperature with respect to the reference temperature" to capture temperature stress levels.

[0070] For categorical variables, one-hot encoding or entity embedding techniques are used to transform them into low-dimensional dense vectors. After the above processing, all static operating condition information is compressed into a fixed-length operating condition feature vector, which completely preserves the external environment and preset conditions at the time of the test.

[0071] The raw waveform data associated with the indicator parameters is used as input. Typically, a single double-pulse test stores two key waveforms in the oscilloscope: the voltage and current waveforms during the turn-on transient of the second pulse, and the voltage and current waveforms during the turn-off transient of the second pulse. The feature extraction algorithm performs the following operations on each waveform segment:

[0072] Automatic identification of critical intervals: By setting voltage thresholds (such as 10% and 90% amplitude) or derivative extreme points, the start and end times of the switching transient and the start and end times of the Miller plateau are automatically located.

[0073] Time-domain feature calculation: Within the identified interval, calculate the peak, mean, and variance of the voltage change rate dv / dt, the peak and mean of the current change rate di / dt, the duration of the Miller plateau, the amplitude of the turn-off voltage overshoot, and the ringing decay time constant.

[0074] Frequency domain characteristic analysis: Fast Fourier Transform (FFT) is applied to the ringing segment of the turn-off voltage to extract the ratio of its main resonant frequency to harmonic amplitude, which is used to characterize the distribution characteristics of parasitic parameters in the power circuit.

[0075] Vector concatenation: All the time-domain, frequency-domain, and morphological features mentioned above are concatenated in a fixed order to form a high-dimensional instantaneous feature vector. This vector is no longer a simple loss value, but a complete dynamic profile of the switching behavior on a sub-microsecond scale.

[0076] Finally, the working condition feature vector and the instantaneous feature vector are stored in parallel in the training sample library as two independent and complementary feature sets. During subsequent model training, they can be input into different branches of the model separately, or concatenated and input into a unified encoder network.

[0077] In one example, time-domain statistical feature extraction is performed on the operating parameters to obtain statistical features; time series analysis is performed on the operating parameters to extract their dynamic change trends to obtain dynamic features; and the statistical features and dynamic features are fused to determine the operating feature vector.

[0078] Among them, statistical features characterize at least one of the mean, variance, and peak value of the operating parameters; dynamic features characterize the slope or frequency component of the operating parameters as they evolve over time.

[0079] For example, the system defines an observation time window that includes the complete switching transient and steady-state current establishment process, starting from the trigger time of the first pulse in the double-pulse test and ending at the completion time of the second pulse turn-off.

[0080] For each operating condition parameter time series recorded at a high sampling rate within this window, a pre-defined set of statistics is calculated channel by channel. Specifically, this includes: calculating the arithmetic mean of all sampling points in the series to reflect the overall intensity level of the operating condition; calculating the unbiased standard deviation of the series to reflect the stability of the operating condition during the test; and scanning the series to find the maximum and minimum values ​​to capture instantaneous drops or overshoot extremes caused by parasitic effects or electromagnetic interference during switching transients. In addition, skewness and kurtosis can be calculated to describe the asymmetry and heavy-tailed characteristics of the distribution. These statistics collectively constitute the statistical characteristic sub-vector of the operating condition parameter.

[0081] Because the duration of double-pulse tests is extremely short (typically several microseconds to tens of microseconds), directly applying traditional trend decomposition methods has limited significance. Therefore, an analysis method optimized for transient processes is adopted. The program performs the following operations on the sequence of operating parameters:

[0082] Differential slope calculation: Perform first-order difference operation on the sequence to obtain the instantaneous rate of change sequence, and then calculate the mean of the rate of change or perform linear regression within the key interval to extract the slope feature, which is used to characterize the rate of current establishment or the rate of voltage drop recovery.

[0083] High-frequency oscillation frequency extraction: The ringing attenuation range after the switch action is intercepted, and the modal parameters are identified by applying Fast Fourier Transform or a feature system-based algorithm. Several frequency components with the largest amplitude and their corresponding attenuation factors are extracted to characterize the characteristics of the bus parasitic inductance and capacitor resonant circuit.

[0084] Envelope characteristics: The upper and lower envelopes of the sequence are extracted by Hilbert transform, and the decay time constant or envelope area of ​​the envelope is calculated as a dynamic feature describing the energy dissipation rate of the transient process.

[0085] The statistical feature vector and the dynamic feature vector obtained after processing the same operating condition parameters through the above two methods are concatenated along the feature dimension to form a joint vector with expanded dimension.

[0086] Optionally, to avoid the curse of dimensionality and feature redundancy, principal component analysis or an autoencoder can be selectively applied to reduce the dimensionality of the concatenated high-dimensional vector and extract the most discriminative latent variables. The final output low-dimensional dense vector is the operating feature vector of the inverter under the current dual-pulse test conditions. This vector is persistently stored and used as one of the input features of the test parameter generation model.

[0087] S203. The working condition feature vector and the instantaneous feature vector are fused to obtain the fused feature vector; the fused feature vector is normalized to obtain the feature vector.

[0088] For example, the working condition feature vector and the instantaneous feature vector are fused; optionally, the fusion operation typically employs one or more of the following strategies:

[0089] Vector concatenation: This method directly concatenates the condition feature vector and the instantaneous feature vector end-to-end along the feature dimension to form a new long vector. This is the most basic fusion method, preserving all the original information, but it may introduce dimensional redundancy.

[0090] Bilinear pooling fusion: Calculate the outer product matrix of the condition feature vector and the instantaneous feature vector, and then compress the outer product matrix into a compact fusion vector using a trainable weight tensor. This method can explicitly model the second-order interaction between static conditions and dynamic behavior.

[0091] Gated attention fusion: Attention weight vectors are calculated separately for the operating condition feature vector and the instantaneous feature vector. The original features are then weighted element-wise before being added or concatenated. This mechanism allows the model to autonomously determine, based on the specific characteristics of the current sample, whether to rely more on static operating point information or on transient waveform details to generate test parameters.

[0092] The output of the fusion operation is a fused feature vector, which now contains information from multiple sources, but the numerical ranges of its various dimensions vary greatly.

[0093] The fused feature vectors are normalized to obtain the feature vectors.

[0094] Optionally, batch normalization or layer normalization can be used as the first layer of built-in operation of the model, and normalization can be completed directly during the forward propagation of the model.

[0095] Optionally, Z-score normalization (x' = (x-μ) / σ) or max-min normalization can be used in the data preprocessing stage to complete the normalization process.

[0096] Iterate through the fused feature vectors of all samples in the training dataset, calculate the mean, standard deviation, minimum, and maximum values ​​for each feature dimension, and persist these statistics as a normalization configuration file. For each dimension of each fused feature vector, apply the selected normalization formula one by one to generate the final feature vector.

[0097] The normalized feature vector has values ​​in all dimensions that are within a similar dynamic range. This vector is then encapsulated as part of the training batch and fed into the input layer of the initial intelligent model.

[0098] In one example, an attention mechanism is used to determine the first weight of each feature element in the working condition feature vector and the second weight of each feature element in the instantaneous feature vector. The working condition feature vector is weighted according to the first weight to obtain a weighted working condition feature vector, and the instantaneous feature vector is weighted according to the second weight to obtain a weighted instantaneous feature vector. The weighted working condition feature vector and the weighted instantaneous feature vector are concatenated to generate a fused feature vector.

[0099] For example, an attention mechanism is used to determine the first weight of each feature element in the working condition feature vector and the second weight of each feature element in the instantaneous feature vector.

[0100] Optionally, this process is implemented using a lightweight, learnable subnetwork. This subnetwork is typically trained end-to-end along with the main body of the test parameter generation model. The specific steps are as follows:

[0101] The working condition feature vector and the instantaneous feature vector are obtained. These can either be used as independent inputs to calculate attention weights separately, or they can be initially concatenated before being fed into a shared attention module. To maintain the independence of the two feature semantic spaces, this embodiment preferably calculates the first weight and the second weight separately.

[0102] For the operating condition feature vector, it is input into the operating condition attention branch, which consists of a fully connected layer and an activation function. This branch first maps the feature vector to a hidden space through a linear transformation, and then outputs a first weight vector with the same dimension as the input vector through a sigmoid or softmax activation function. The value of each element in the first weight vector is between 0 and 1 (if using sigmoid) or the sum is 1 (if using softmax), corresponding to the relative importance score of each element in the original operating condition feature vector. Similarly, the instantaneous feature vector is passed through the instantaneous attention branch to generate a second weight vector.

[0103] The parameters of the attention subnetwork are continuously optimized through backpropagation during training. This means that when the model finds that certain static operating condition features are crucial for predicting switching delays under high-voltage conditions, the weight connection corresponding to that feature will receive a large gradient update, so that when the model encounters high-voltage samples, the first weight corresponding to the junction temperature dimension approaches 1; conversely, for features with negligible impact, their weights will be trained to approach 0.

[0104] The program multiplies the original working condition feature vector element-wise with the calculated first weight vector to obtain the weighted working condition feature vector. Mathematically, this operation is equivalent to performing a coordinate axis scaling transformation on the original feature space based on task importance. Similarly, the original instantaneous feature vector is multiplied element-wise with the second weight vector to obtain the weighted instantaneous feature vector.

[0105] The weighted working condition feature vector and the weighted instantaneous feature vector are concatenated end-to-end along the feature dimension to generate a fused feature vector that contains both weighted static and weighted dynamic information. This vector is then fed into the normalization layer and the subsequent backbone network for further processing.

[0106] Optionally, during the model inference and deployment phase, the weight parameters of the attention module are frozen. When a new test sample is input, the attention branch automatically calculates a weight vector that adapts to the current operating conditions and transient waveforms within a very short forward propagation time, based on the learned rules, thus completing dynamic feature weighted fusion.

[0107] S204. Input the feature vector into the initial intelligent model, process the feature vector based on the initial intelligent model to obtain the initial test parameters; determine the target value based on the initial test parameters, the test parameter labels of the feature vector, and the weighted objective function; optimize the parameters in the initial intelligent model according to the target value using the chain rule to obtain the test parameter generation model.

[0108] For example, the feature vector is associated with a corresponding test parameter label, which is a test parameter reference value that corresponds one-to-one with the operating condition parameter and the index parameter.

[0109] For example, the feature vectors that have been previously fused and normalized are fed into the input layer of the initial intelligent model in batches. The initial intelligent model is composed of multiple layers of linear transformation matrices and nonlinear activation functions stacked together.

[0110] The feature vector flows sequentially through these computational layers: each neuron receives a weighted sum of the outputs of the previous layer, adds a bias term, and then introduces non-linear expressive power through activation functions such as ReLU, Sigmoid, or GELU. After successive abstraction and transformation through all hidden layers, a set of numerical vectors is finally generated in the output layer, which are the initial test parameters. In the first training epoch, since the model parameters are randomly initialized, this output vector exhibits a significant random error distribution compared to the true test parameter labels.

[0111] The initial test parameter vector output by the model, along with the corresponding test parameter label vector read from the dataset, are input into a predefined weighted objective function. This function first calculates the error between the predicted value and the label value for each dimension of the test parameter. Commonly used basic error metrics include mean squared error, mean absolute error, or Huber loss. Then, the program reads a pre-defined weight vector and performs a weighted summation of the error terms for each dimension. The resulting single scalar value is the objective value. The magnitude of this objective value directly reflects the quality of the parameters generated by the current model—the lower the objective value, the closer the model's output test parameters are to historically validated successful settings.

[0112] Subsequently, backpropagation and parameter optimization based on the chain rule are performed. The optimizer (such as Adam or SGD) takes over the control flow. Starting from the scalar node of the target value, the automatic differentiation engine recursively applies the chain rule in the reverse direction of the computation graph (i.e., from the output layer to the input layer), calculating the partial derivative of the target value with respect to each trainable parameter, i.e., the gradient. The gradient vector indicates the direction in which the objective function value increases the fastest under the current parameter state. To minimize the target value, the optimizer fine-tunes and updates the model parameters along the reverse direction of the gradient; the update magnitude is controlled by the learning rate hyperparameter. The update formula is typically: New parameters = Old parameters - Learning rate × Gradient.

[0113] A complete training iteration consists of the following steps: forward propagation to calculate the initial test parameters, comparison with labels to calculate the target value, backpropagation to calculate the gradient, and finally updating the model parameters. This iteration is repeated hundreds to thousands of times on the training dataset. As the iteration progresses, the target value gradually converges to a very low steady-state level, indicating that the weighted error between the initial test parameters and the test parameter labels has been minimized. At this point, the program saves the model's network structure definition and the final optimized parameter weight file. This file and the loaded running instance constitute the test parameter generation model.

[0114] In one example, the predicted switching loss is determined based on the initial test parameters and the test parameter labels in the feature vector; the switching loss error is determined based on the predicted switching loss and the switching loss in the feature vector; the index oscillation amplitude is determined based on the index parameters in the feature vector and the preset index threshold; and the weighted objective function is determined with the goal of minimizing the switching loss error and minimizing the index oscillation amplitude.

[0115] For example, after the initial intelligent model outputs a set of initial test parameters, the model does not simply compare these parameters with the labels and call it a day. Instead, it feeds them into a differentiable loss estimation proxy module. This proxy module can be a simplified formula based on the device's physical equations, or it can be a pre-trained micro neural network.

[0116] In one possible implementation, the predicted value of the switching loss can be calculated according to the formula E_off_pred = k×V_bus×I_load×(Rg_off)^α×(T2)^β.

[0117] Where E_off_pred is the predicted switching loss; k is the preset proportional constant; V_bus is the DC bus voltage; I_load is the load current; Rg_off is the gate turn-off resistor; α is the turn-off resistor exponent; T2 is the second pulse width; and β is the pulse width exponent.

[0118] The model compares the predicted loss value with the actual measured loss labels stored in the feature vector to calculate the switching loss error term. The existence of this error forces the model to respect the actual energy dissipation characteristics of the device under the current operating conditions when generating test parameters, avoiding the generation of a set of invalid parameters that can establish current but distort the loss measurement results.

[0119] The model reads a subset of instantaneous features from the feature vectors, particularly components related to dynamic stress. Simultaneously, it reads a pre-defined table of preset threshold values. Oscillation amplitude is typically calculated using a hinge loss function or a quadratic penalty term for values ​​exceeding the threshold.

[0120] In one possible implementation, if the voltage overshoot threshold is set to V_th = 1.2, and the parameters generated by the current model are expected to cause the overshoot ratio to reach V_os, then the index oscillation amplitude term L_osc is defined as: L_osc = max(0, V_os - V_th)². This means that as long as the predicted electrical stress is within the safety threshold, this loss is zero, and the model is not penalized; once the predicted stress exceeds the threshold, the loss value increases rapidly with the square of the excess, forming a strong rejection signal.

[0121] The two error components mentioned above are linearly combined with their corresponding weighting coefficients to form the final weighted objective function. This function is optimized by minimizing both switching loss error and index oscillation amplitude. Its mathematical expression is: Total_Loss=λ_E×L_E+λ_osc×L_osc.

[0122] Where Total_Loss is the total target loss value; λ_E is the switching loss error weight coefficient, usually set to 1.0; L_E is the switching loss error; λ_osc is the safety constraint weight, since safety takes priority over measurement accuracy, this value is usually set between 5.0 and 20.0 to ensure that any small breach of the safety boundary will generate a dominant gradient signal; L_osc is the index oscillation amplitude term.

[0123] In each training iteration, Total_Loss serves as the target value, driving the backpropagation of the subsequent chain rule.

[0124] This application provides a training method for a test parameter generation model. By acquiring the inverter's operating condition parameters and index parameters, feature extraction is performed on the operating condition parameters to obtain operating condition feature vectors, and instantaneous feature extraction is performed on the index parameters to obtain instantaneous feature vectors. The two types of feature vectors are then fused and normalized to obtain standardized feature vectors. Subsequently, the feature vectors are input into an initial intelligent model to generate initial test parameters. The target value is determined by combining the initial test parameters, the test parameter labels corresponding to the feature vectors, and a weighted objective function. Finally, the model parameters are iteratively optimized using a chain rule. This fully explores the intrinsic correlation between the inverter's operating conditions and test indicators, achieving accurate fitting of the model to the parameter mapping relationship. It avoids the limitations of manual feature engineering and achieves the effect of quickly and accurately generating dual-pulse test parameters adapted to different inverter operating conditions. This significantly reduces the dependence of test parameter generation on human experience, significantly improves parameter generation efficiency, adaptation accuracy, and generalization ability, and ensures the stability and reliability of inverter dual-pulse testing.

[0125] Figure 3 A flowchart illustrating an optimized method for dual-pulse testing provided in this application embodiment is shown below. Figure 3 As shown, the method includes:

[0126] S301. Obtain the inverter's operating parameters.

[0127] For example, the operating parameters of the inverter are obtained. The test system or controller establishes a connection with the inverter's main control unit through the communication interface and sends a parameter query command. Next, the inverter firmware responds to the query, reading the pre-stored device identifier, nameplate ratings, and hardware configuration list from its non-volatile memory. Then, the system parses and verifies the returned data frame, extracting the operating parameter fields: obtaining the inverter model string from the device description block; reading the rated DC voltage and rated AC current RMS values ​​from the electrical parameter register; obtaining the IGBT model code from the power module identification register or directly reading device specification parameters, such as collector-emitter breakdown voltage, nominal collector current, and typical saturation voltage drop.

[0128] Finally, the system integrates the aforementioned discrete parameter values ​​into a structured operating condition parameter dataset, stores it in a memory buffer, and uses it as input for subsequent double-pulse testing, including bus voltage preset, pulse width calculation, and protection threshold setting. If communication fails or parameters are missing, an exception handling process is triggered, prompting manual input or the use of default conservative values.

[0129] S302. Based on the operating condition parameters, perform matching processing in the preset database to determine the initial test parameters.

[0130] For example, the system receives operating condition parameters; optionally, the operating condition parameters include, but are not limited to, inverter model, rated voltage, rated current, and IGBT specification fields.

[0131] The system invokes the database query engine and performs a search according to a preset matching priority strategy. The first priority is exact model matching, which uses the inverter model string as the keyword to search for a completely matching record in the database's model index table. If a match is found, all associated initial test parameters are directly read. If the first priority fails, the system switches to the second priority, which is fuzzy matching based on IGBT specifications. The system extracts the voltage and current ratings from the IGBT specifications and searches for entries with the same voltage rating and closest current rating in the device parameter mapping table. The third priority is matching based on the ratio of rated voltage to current. If the first two priorities fail, the system calculates and generates a conservative set of initial parameters based on the percentage of rated voltage and rated current.

[0132] The matched parameter values ​​will undergo a rationality check, such as checking whether the voltage value exceeds 80% of the IGBT withstand voltage or the upper limit of the test equipment range. After the check is passed, the set of parameters is determined as the initial test parameters for this test and loaded into the test sequence generator.

[0133] S303. Perform double-pulse testing on the inverter based on the operating condition parameters and initial test parameters to obtain the initial index parameters; generate a model based on the operating condition parameters, initial index parameters, and test parameters to determine the target index parameters corresponding to the operating condition parameters.

[0134] For example, the test parameter generation model is trained using any of the training methods described in the above embodiments.

[0135] For example, after completing the first round of double-pulse testing and obtaining initial performance parameters, the system calls the built-in test parameter generation model. This model receives a two-dimensional input vector: the first vector is the fixed operating condition parameters; the second vector is the initial performance parameters obtained from the previous round of testing, such as the actual turn-off voltage spike, switching loss value, and current rise rate.

[0136] The model performs calculations based on a preset control strategy. For example, if the ratio of the measured turn-off voltage spike to the device withstand voltage is too high, it outputs a command to reduce the bus voltage; if the measured current does not reach the set ratio of the rated current, it outputs a command to increase the first pulse width.

[0137] The model calculation results are output in the form of optimized test parameters and directly updated to the register of the test sequence generator, triggering a new round of dual-pulse hardware timing control.

[0138] The system performs a second or Nth double-pulse test on the inverter based on the new optimized test parameters, re-acquires waveforms, and calculates new performance parameters. The system compares the newly acquired performance parameters with convergence criteria, which can be: the rate of change of switching losses between two consecutive tests is less than a preset threshold, or the deviation between the measured current value and the target current value falls within the allowable error band. If the convergence criteria are met, the performance parameters for that round are determined as the final target performance parameters, and the loop exits. If not, the operating parameters and the new performance parameters are input into the model again for iteration until the conditions are met or the preset maximum number of iterations is reached.

[0139] In one example, the following steps are repeated until a preset condition is met: Input the operating condition parameters and the i-th initial index parameter into the test parameter generation model to obtain the i-th test parameter output by the test parameter generation model; perform a double-pulse test on the inverter based on the operating condition parameters and the i-th test parameter to obtain the (i+1)-th initial index parameter; where the first initial index parameter is the initial index parameter; i is a positive integer greater than or equal to 1; and the value of i is incremented by 1; the initial index parameter obtained when the preset condition is met is used to determine the target index parameter corresponding to the operating condition parameters.

[0140] For example, the system initializes the iteration counter, setting i=1. At this time, the first initial index parameter obtained in the first round of testing already exists in memory, that is, the initial index parameter mentioned above.

[0141] Entering the loop: The system uses the unchanged operating parameters and the latest i-th initial index parameter as a joint input vector, and submits it to the test parameter generation model for forward inference. The model calculates the adjustment direction and magnitude according to the built-in algorithm logic, generates and outputs the i-th test parameter. This parameter is then loaded into the hardware control unit, triggering a new round of double-pulse test processing on the inverter.

[0142] After the test is completed, the waveform analysis module calculates the new dynamic performance value and stores it as the (i+1)th initial index parameter. Next, the system performs an increment operation, setting i = i+1, completing one full iteration cycle. Then, the system calls the preset condition judgment module to evaluate the current state. If the preset condition is not met, the program flow jumps back to the beginning of the loop, and the above process is executed again with the newly generated index parameter as input. If the preset condition is met, the loop immediately exits. Finally, the system formats and extracts the set of initial index parameters corresponding to the loop exit, defines it as the final output target index parameter under that operating condition, and writes it into the test report.

[0143] Understandably, the preset conditions can be one or more of the following: the relative deviation between two consecutive loss values ​​is greater than 5%; or, i is less than the maximum number of iterations, 10.

[0144] It is understood that after completing the aforementioned iterative optimization and parameter determination steps, the embodiments of this application further cover the data archiving, multi-dimensional automatic traversal, and final result output stages.

[0145] During the target parameter locking phase, the system automatically records the optimal parameter combination that has been converged and verified under the current specific operating condition. This includes key control variables such as the conduction pulse width setting of the dual pulse, the drive voltage amplitude configuration, and the dead time of the bridge arm commutation. The system then associates this set of parameters with the corresponding operating condition label and measured indicators and stores them in the background database. This not only solidifies the test results but also provides new real data samples for the test parameter generation model, supporting the continuous online learning and accuracy evolution of the model algorithm.

[0146] During the full-condition coverage test phase, the system automatically executes the condition traversal logic based on the preset test matrix. Through the linkage control of the programmable power supply, temperature control box and electronic load, it autonomously switches to different test points. For example, the output load current is set to 25%, 50% and 100% of the rated value in sequence, and the ambient temperature is adjusted to the low temperature, normal temperature and high temperature threshold. Under each switched condition node, the system will repeatedly call the aforementioned iterative test and parameter optimization steps until all condition combinations in the matrix are traversed. Finally, a parameter mapping table that fully covers the expected operating range of the inverter is formed. This mapping table fully reveals the multi-dimensional relationship between pulse width, dead time and switching performance under different stress conditions.

[0147] In the test report generation phase, the system summarizes all raw waveform data and calculated indicators generated during the full-condition testing phase and generates a formatted formal test document through an automated reporting engine. The report not only provides detailed outputs of turn-on losses, turn-off losses, and dynamic switching trajectory curves under various operating conditions in the form of charts and numerical values, but also innovatively introduces a waveform quality scoring mechanism. This mechanism quantifies and scores the measured waveforms based on preset indicators such as overshoot suppression rate, oscillation decay time, and Miller plateau flatness. It also compares and labels the actual waveform scores under the initially recommended test parameters and the final locked parameters, thus providing R&D engineers with detailed and reliable data to intuitively evaluate the dynamic characteristics of the inverter and the effectiveness of the recommended test model.

[0148] This application provides an optimized method for double-pulse testing. It acquires the inverter's operating parameters, performs matching processing in a pre-set database based on these parameters to determine initial test parameters, and then conducts double-pulse testing on the inverter using both the operating parameters and the initial test parameters to obtain corresponding initial index parameters. Finally, it combines the operating parameters, initial index parameters, and a pre-trained test parameter generation model to accurately lock the target index parameters corresponding to the operating parameters. This closed-loop process achieves rapid initial screening of test parameters through database matching, avoiding the blindness and repeated trial-and-error costs of manual parameter setting. Furthermore, it combines measured data with a dedicated generation model to achieve precise matching of parameters and indexes, fully aligning with the real-time operating characteristics of the inverter. This achieves intelligent and efficient optimization of the entire double-pulse testing process, significantly shortening the test parameter debugging cycle, significantly improving the operating condition adaptation accuracy of test indicators and the reliability of test results, and effectively reducing the excessive reliance on manual experience in double-pulse testing.

[0149] Figure 4 This is a schematic diagram of the structure of a training device for a test parameter generation model provided in an embodiment of this application, as shown below. Figure 4 As shown, the training device 40 for a test parameter generation model provided in this embodiment includes:

[0150] The first acquisition module 401 is used to acquire the operating condition parameters and index parameters of the inverter; wherein, the operating condition parameters and index parameters are paired data with a one-to-one correspondence acquired under the same double-pulse test conditions.

[0151] The first processing module 402 is used to perform feature extraction processing on the operating condition parameters and index parameters to obtain feature vectors.

[0152] The training module 403 is used to input feature vectors into the initial intelligent model, process the feature vectors based on the initial intelligent model, and optimize the parameters of the initial intelligent model through a weighted objective function to obtain the test parameter generation model. The test parameter generation model is used to process the inverter's operating parameters and index parameters to obtain test parameters. The test parameters are used to complete the double-pulse test of the inverter.

[0153] In one possible implementation, the first processing module 402 includes:

[0154] The processing submodule 4021 is used to perform feature extraction processing on the operating condition parameters to obtain the operating condition feature vector; and to perform instantaneous feature extraction processing on the index parameters to obtain the instantaneous feature vector. Among them, the operating condition feature vector represents the static operating state and working conditions of the inverter under the double pulse test; and the instantaneous feature vector represents the transient change characteristics of the electrical parameters of the inverter during the double pulse test.

[0155] The fusion module 4022 is used to fuse the working condition feature vector and the instantaneous feature vector to obtain the fused feature vector; and to normalize the fused feature vector to obtain the feature vector.

[0156] In one possible implementation, the processing submodule 4021 includes:

[0157] The operating parameters are subjected to time-domain statistical feature extraction to obtain statistical features; the operating parameters are subjected to time series analysis to extract their dynamic change trend to obtain dynamic features; among them, the statistical features represent at least one of the mean, variance, and peak value of the operating parameters; the dynamic features represent the slope or frequency component of the operating parameters as they evolve over time.

[0158] The statistical and dynamic features are fused to determine the operating condition feature vector.

[0159] In one possible implementation, the fusion module 4022 includes:

[0160] An attention mechanism is used to determine the first weight of each feature element in the working condition feature vector and the second weight of each feature element in the instantaneous feature vector.

[0161] The working condition feature vector is weighted according to the first weight to obtain the weighted working condition feature vector, and the instantaneous feature vector is weighted according to the second weight to obtain the weighted instantaneous feature vector.

[0162] The weighted working condition feature vector and the weighted instantaneous feature vector are concatenated to generate a fused feature vector.

[0163] In one possible implementation, the feature vector is associated with a corresponding test parameter label, and the test parameter label is a test parameter reference value that corresponds one-to-one with the operating condition parameter and the index parameter; the training module 403 includes:

[0164] The feature vectors are input into the initial intelligent model, and the feature vectors are processed based on the initial intelligent model to obtain the initial test parameters.

[0165] The target value is determined based on the initial test parameters, the test parameter labels of the feature vector, and the weighted objective function.

[0166] Based on the target value, the parameters in the initial intelligent model are optimized using the chain rule to obtain the test parameter generation model.

[0167] In one possible implementation, the feature vector is associated with a corresponding test parameter label, and the test parameter label is a test parameter reference value that corresponds one-to-one with the operating condition parameter and the index parameter; the training device 40 also includes:

[0168] Based on the initial test parameters and the test parameter labels in the feature vector, the predicted switching loss is determined; based on the predicted switching loss and the switching loss in the feature vector, the switching loss error is determined.

[0169] The oscillation amplitude of the indicator is determined based on the indicator parameters in the feature vector and the preset indicator threshold.

[0170] The weighted objective function is determined with the goals of minimizing switching loss error and minimizing index oscillation amplitude.

[0171] This embodiment provides a training device for a test parameter generation model, which can execute the method provided in the above-described method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0172] Figure 5 This is a schematic diagram of the structure of an optimized device for dual-pulse testing provided in an embodiment of this application, as shown below. Figure 5 As shown, the optimized device 50 for dual-pulse testing provided in this embodiment includes:

[0173] The acquisition module 501 is used to acquire the operating parameters of the inverter; based on the operating parameters, it performs matching processing in a preset database to determine the initial test parameters.

[0174] The processing module 502 is used to perform double-pulse test processing on the inverter based on the operating condition parameters and the initial test parameters to obtain the initial index parameters.

[0175] The determination module 503 is used to generate a model based on the operating condition parameters, initial index parameters, and test parameters, and to determine the target index parameters corresponding to the operating condition parameters; the test parameter generation model is trained by the third aspect and / or various possible training devices mentioned above.

[0176] In one possible implementation, the determining module 503 includes:

[0177] Repeat the following steps until the preset conditions are met:

[0178] Input the operating condition parameters and the i-th initial index parameter into the test parameter generation model to obtain the i-th test parameter output by the test parameter generation model.

[0179] Based on the operating parameters and the i-th test parameter, the inverter is subjected to a double-pulse test to obtain the (i+1)-th initial index parameter; where the first initial index parameter is the initial index parameter; i is a positive integer greater than or equal to 1; and the value of i is determined to be incremented by 1.

[0180] Among them, the initial index parameters obtained when the preset conditions are met are used to determine the target index parameters corresponding to the operating condition parameters.

[0181] This embodiment provides an optimized device for dual-pulse testing, which can execute the method provided in the above-described method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0182] This embodiment provides an inverter with an integrated test module, comprising:

[0183] An inverter is used to convert direct current (DC) to alternating current (AC).

[0184] The test module, electrically connected to the inverter, is used to execute the training method of the test parameter generation model in the above embodiments, or the optimization method of the double-pulse test in the above embodiments.

[0185] The testing module includes: a memory for storing operating condition parameters, index parameters, and test parameter generation models; and a processor for executing training operations of the test parameter generation models or optimization operations of double-pulse tests.

[0186] This embodiment provides an inverter with an integrated testing module, which can execute the methods provided in the above-described method embodiments. Its implementation principle and technical effects are similar, and will not be described in detail here.

[0187] Figure 6This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device 60 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the electronic device 60 further includes a communication component 603. The processor 601, memory 602, and communication component 603 are connected via a bus 604.

[0188] In a specific implementation, at least one processor 601 executes computer execution instructions stored in memory 602, causing at least one processor 601 to perform the above-described method.

[0189] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0190] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0191] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0192] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0193] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0194] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0195] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0196] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0197] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0198] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0199] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0200] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0201] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0202] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A training method for a test parameter generation model, characterized in that, include: Obtain the inverter's operating condition parameters and index parameters; wherein the operating condition parameters and index parameters are paired data with a one-to-one correspondence obtained under the same double-pulse test conditions; The operating condition parameters and the index parameters are subjected to feature extraction processing to obtain feature vectors; wherein, the feature vectors are associated with corresponding test parameter labels, and the test parameter labels are test parameter reference values ​​that correspond one-to-one with the operating condition parameters and the index parameters; The feature vector is input into an initial intelligent model, the feature vector is processed based on the initial intelligent model, and the parameters of the initial intelligent model are optimized through a weighted objective function to obtain a test parameter generation model, including: The feature vector is input into the initial intelligent model, and the feature vector is processed based on the initial intelligent model to obtain the initial test parameters; Based on the initial test parameters, the test parameter labels of the feature vector, and the weighted objective function, the target value is determined; Based on the target value, the parameters in the initial intelligent model are optimized using a chain rule to obtain the test parameter generation model; The test parameter generation model is used to process the inverter's operating parameters and index parameters to obtain test parameters; the test parameters are used to complete the double-pulse test of the inverter.

2. The method according to claim 1, characterized in that, The operating condition parameters and the index parameters are subjected to feature extraction processing to obtain a feature vector, including: The operating condition parameters are subjected to feature extraction processing to obtain an operating condition feature vector; the index parameters are subjected to instantaneous feature extraction processing to obtain an instantaneous feature vector; wherein, the operating condition feature vector represents the static operating state and working conditions of the inverter under the double pulse test; and the instantaneous feature vector represents the transient change characteristics of the electrical parameters of the inverter during the double pulse test. The working condition feature vector and the instantaneous feature vector are fused to obtain a fused feature vector; the fused feature vector is then normalized to obtain the feature vector.

3. The method according to claim 2, characterized in that, The operating condition parameters are subjected to feature extraction processing to obtain an operating condition feature vector, including: The operating parameters are subjected to time-domain statistical feature extraction to obtain statistical features; the operating parameters are subjected to time series analysis to extract their dynamic change trend to obtain dynamic features; wherein, the statistical features represent at least one of the mean, variance, and peak value of the operating parameters; the dynamic features represent the slope or frequency component of the evolution of the operating parameters over time. The statistical features and the dynamic features are fused to determine the operating condition feature vector.

4. The method according to claim 2, characterized in that, The condition feature vector and the instantaneous feature vector are fused to obtain a fused feature vector, including: An attention mechanism is used to determine the first weight of each feature element in the working condition feature vector and the second weight of each feature element in the instantaneous feature vector. The working condition feature vector is weighted according to the first weight to obtain a weighted working condition feature vector, and the instantaneous feature vector is weighted according to the second weight to obtain a weighted instantaneous feature vector. The weighted working condition feature vector and the weighted instantaneous feature vector are concatenated to generate the fused feature vector.

5. The method according to any one of claims 1-4, characterized in that, The feature vector is associated with a corresponding test parameter label, and the test parameter label is a test parameter reference value that corresponds one-to-one with the operating condition parameter and the index parameter; the method further includes: Based on the initial test parameters and the test parameter labels in the feature vector, the predicted switching loss is determined; based on the predicted switching loss and the switching loss in the feature vector, the switching loss error is determined. The oscillation amplitude of the indicator is determined based on the indicator parameters in the feature vector and the preset indicator threshold. The weighted objective function is determined with the goals of minimizing the switching loss error and minimizing the index oscillation amplitude.

6. An optimized method for double-pulse testing, characterized in that, include: Obtain the inverter's operating parameters; perform matching processing on the operating parameters in a preset database to determine the initial test parameters; Based on the operating parameters and the initial test parameters, the inverter is subjected to a double-pulse test to obtain the initial index parameters. Based on the operating condition parameters, the initial index parameters, and the test parameter generation model, the target index parameters corresponding to the operating condition parameters are determined; wherein, the test parameter generation model is trained by the training method for the test parameter generation model as described in any one of claims 1-5.

7. The method according to claim 6, characterized in that, Based on the operating condition parameters, the initial index parameters, and the test parameter generation model, the target index parameters corresponding to the operating condition parameters are determined, including: Repeat the following steps until the preset conditions are met: The operating condition parameters and the i-th initial index parameter are input into the test parameter generation model to obtain the i-th test parameter output by the test parameter generation model; Based on the operating parameters and the i-th test parameter, the inverter is subjected to a double-pulse test to obtain the (i+1)-th initial index parameter; where the first initial index parameter is the initial index parameter; i is a positive integer greater than or equal to 1; and the value of i is determined to be incremented by 1. The initial index parameters obtained when the preset conditions are met are used to determine the target index parameters corresponding to the operating condition parameters.

8. An optimized device for dual-pulse testing, characterized in that, include: The acquisition module is used to acquire the operating condition parameters of the inverter; and to perform matching processing on the operating condition parameters in a preset database to determine the initial test parameters. The processing module is used to perform a double-pulse test on the inverter based on the operating condition parameters and the initial test parameters to obtain the initial index parameters; The determination module is used to determine the target index parameter corresponding to the operating condition parameter based on the operating condition parameter, the initial index parameter, and the test parameter generation model; wherein the test parameter generation model is trained by the training method of the test parameter generation model as described in any one of claims 1-5.

9. An inverter with an integrated test module, characterized in that, include: An inverter is used to convert direct current (DC) to alternating current (AC). The test module, electrically connected to the inverter, is used to execute the training method of the test parameter generation model as described in any one of claims 1-5, or the optimization method of the double pulse test as described in any one of claims 6-7; The testing module includes: a memory for storing operating condition parameters, index parameters, and a test parameter generation model; and a processor for executing the training of the test parameter generation model or the optimization operation of the double-pulse test.

Images