Performance evaluation method of MOS tube for electronic load based on multi-feature pattern recognition

Abstract

The present application proposes a performance evaluation method for MOS tubes for electronic loads based on multi-feature pattern recognition. The method includes high-frequency acquisition of the sampled voltage of the load circuit and the drain voltage of each MOS tube, and sliding slices to form a sample set after preprocessing the sampled signal; extracting 4 features from the sliced ​​samples: ① the original sample data of the sliding slice, ② performing spectrum transformation, and forming a "frequency-spectrum value-threshold" pair with the 10 highest spectrum values ​​and the corresponding frequency, ③ establishing power-on and power-off curve description parameters to form a curve morphology parameter combination, and ④ drawing cyclic dual-spectrum graphic features; establishing a Siamese small sample model to predict the single-feature state of the circuit; using the single-feature state as a gene fragment, and using a genetic algorithm to mine the MOS tube performance classification evaluation model. The present application provides a method for performance evaluation of MOS tubes for electronic loads, screening out unqualified MOS tubes, and ensuring the health and stability of the electronic load load circuit.

Description

Technical Field

[0001] The present invention belongs to the field of big data driven deep network model hardware circuit performance evaluation, and specifically relates to a MOS tube performance evaluation method for electronic loads based on multi-feature pattern recognition. Background Art

[0002] MOS tubes are the main components of electronic loads. MOS tubes can cause tube burning due to performance defects. The maximum current of an electronic load can reach 3000A or even higher, requiring hundreds of MOS tubes to run in parallel, and the voltage of a series branch can reach 2000V or even higher. If a MOS tube is delayed in conduction in a series branch of a parallel circuit, no current will flow through this series branch at the moment of power-on, which will cause the current of other parallel branches to increase. If the circuit design redundancy is not enough, all parallel branch MOS tubes will be burned, causing great economic losses. On the contrary, if all load circuits have sufficient redundancy, the circuit cost will be greatly increased. Therefore, before the electronic load circuit leaves the factory, the high-frequency voltage data of the key position of the circuit is collected, and the performance of each MOS tube is evaluated based on a deep learning method driven by big data, which has important practical significance for improving the economic benefits of the enterprise. The MOS tube performance evaluation method based on multi-feature pattern recognition proposed in this application can quickly and accurately screen out sub-healthy MOS tubes with performance defects.

[0003] At present, the mainstream performance evaluation method is mainly based on hardware experiments, which is too dependent on manual work and requires a lot of manpower and material resources. At present, there is no economical and practical method to solve the high performance requirements of MOS tubes for DC electronic loads. Summary of the invention

[0004] The purpose of the present invention is to provide a performance evaluation method for MOS tubes used in electronic loads, which can select sub-healthy MOS tubes with performance defects to ensure the safe and stable operation of the electronic load.

[0005] In order to solve the above technical problems, the present invention proposes a MOS tube performance evaluation method for electronic load based on multi-feature pattern recognition, comprising:

[0006] The load current sampling voltage and the drain voltage of each MOS tube are collected at high frequency, and a GRU multi-input multi-output timing signal denoising model is established. The pre-processed noise-free sampling data is sliced ​​to form a slice sample set.

[0007] The denoised slice samples are used as the first feature; spectrum transformation is performed, and the "frequency-spectrum value" pairs composed of the 10 highest spectrum values ​​and the corresponding frequencies are used as the second feature; the power-on and power-off curve description parameters are established, and the power-on and power-off morphological parameter combinations formed are used as the third feature; the drawn cyclic bispectrum is used as the fourth feature;

[0008] A Siamese single-feature state prediction model with an EMA attention mechanism module is established to predict the circuit state based on a single feature based on the principle of small sample feature matching;

[0009] The single characteristic state of each circuit is used as a fragment of the genetic algorithm genome, and the MOS tube classification performance is used as a label to mine the classification mode of MOS tube performance evaluation for electronic loads;

[0010] The performance classification and evaluation of MOS tubes in DC load circuits are carried out based on the evaluation model mined by genetic algorithm.

[0011] Optionally, the step of establishing a GRU multi-input multi-output time series signal denoising model and slidingly slicing the preprocessed noise-free sampled data to form a slice sample set includes:

[0012] Based on the least squares objective function, a multi-input and multi-output GRU time series network is trained to denoise the sampled data. The noise-free data is sliced ​​by sliding, and each slice is used as a sample. The samples are divided into training set, validation set and test set.

[0013] Optionally, four features are extracted from the slice denoising sample, including:

[0014] In order to extract the circuit state information contained in the original data, the denoised slice samples are used as the first feature. In order to extract the circuit state information in the frequency domain, the slice samples are spectrally transformed, and the "frequency-spectrum value" pairs consisting of the 10 highest spectrum values ​​and the corresponding frequencies are used as the second feature. The power-on and power-off curve description parameters are established, and the power-on and power-off morphological parameter combination formed is used as the third feature. In order to extract the periodic state information, the cyclic bispectral slice diagram is drawn as the fourth feature.

[0015] Optionally, the step of establishing a Siamese single-feature state prediction model with an EMA attention mechanism module includes:

[0016] Select a one-shot feature matching algorithm, establish a ternary sample set in the form of "state A, same type state P with different working conditions, different type state N", select the ternary center loss function to train the Siamese single feature state prediction model, use the extracted 4 features as input, train the EMA-Siamese model of the 1D or 2D convolutional architecture, and predict the 4 single feature states of the circuit.

[0017] Optionally, mining the MOS tube performance evaluation classification mode for DC electronic load based on genetic algorithm includes:

[0018] The performance of MOS tubes is jointly determined by four single-feature states. In order to mine the MOS tube performance types corresponding to the four-state combination, the single-feature state is used as a fragment of the four-fragment genome of the genetic algorithm. The position and content of the gene fragment are numerically encoded according to the MOS tube performance and the single-feature circuit state. The multi-feature state combination pattern corresponding to the classification performance of MOS tubes is mined through the crossover mutation optimization selection method.

[0019] Optionally, matching the combination of multiple characteristic states with the MOS tube performance evaluation mode and identifying the MOS tube performance type according to the matching result includes:

[0020] Four performance types of "excellent, good, qualified, and poor" are established, and the four characteristics are used as Siamese model input. Four circuit single feature states are given through the ternary feature matching method. The four states are combined to match the MOS tube performance evaluation pattern library, and the performance classification evaluation of the MOS tube is given based on the correspondence between the matching pattern and the MOS tube performance type.

[0021] The present application proposes a performance evaluation method for MOS tubes for electronic loads based on multi-feature pattern recognition. The method includes high-frequency acquisition of the sampled voltage of the load circuit and the drain voltage of each MOS tube, preprocessing the sampled signal and then sliding the slice to form a sample set; extracting 4 features from the sliced ​​sample: ① the original data of the sample of the sliding slice, ② performing spectrum transformation, and forming a "frequency-spectrum value-threshold" pair with the 10 highest spectrum values ​​and the corresponding frequency, ③ establishing the description parameters of the power-on and power-off curves to form a curve morphology parameter combination, ④ drawing the cyclic dual spectrum graphic features; establishing a Siamese small sample model to predict the single feature state of the circuit; using the single feature state as a gene fragment, and using a genetic algorithm to mine the classification mode of the performance evaluation of MOS tubes for electronic loads. The present application provides a method for performance evaluation of MOS tubes for electronic loads, screening unqualified MOS tubes, and ensuring the health and stability of the electronic load load circuit. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] FIG1 is a schematic flow chart of a method for evaluating the performance of a MOS tube for an electronic load based on multi-feature pattern recognition according to an embodiment of the present application;

[0024] FIG2 is a multi-MOS tube series-parallel load circuit provided by an embodiment of the present invention;

[0025] FIG. 3 is a diagram of a genetic algorithm performance evaluation pattern mining framework provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0027] As shown in FIG. 1 , FIG. 1 is a flow chart of a method for evaluating the performance of a MOS tube for an electronic load based on multi-feature pattern recognition provided in an embodiment of the present application, which specifically includes five contents.

[0028] S11: High frequency acquisition of the sampling voltage of the load circuit and the drain voltage of each MOS tube, pre-processing the sampling signal and then sliding the slice to form a sample set.

[0029] It should be noted that the original voltage data collected on the MOS tube of the load circuit is often mixed with high-frequency noise and abnormal outliers that affect the evaluation effect. Therefore, we need to denoise the original sampled data before feature extraction and slide slice it into sample data, which mainly includes two steps.

[0030] S111: Add various types of noise and outliers to the original collected data to form a noisy training sample set, select a multi-input, multi-output GRU time series neural network, use the long sequence of noisy sampled data as the time series input of the model, and the model outputs noise-free time series data. The training model minimizes the point-to-point least squares loss between the model output and the original sampled data. The GRU unit uses the reset gate and the update gate to forget the unimportant feature information of the historical sampling points in the long sequence input and remember the important feature information. xt and the previous unit hidden layer output h t-1 Based on, calculate the update gate z at the current sampling point input t and reset gate r t , as shown in equations (1) and (2),

[0031] (1)

[0032] (2)

[0033] Then, the candidate output of the current sampling point is calculated based on the feature information of the current sampling point and the previous sampling point retained by the reset gate. , as shown in formula (3), where Represents the important characteristic information of the current sampling point.

[0034] (3)

[0035] The new gate selectively memorizes and forgets the feature information of the previous sampling point and the current sampling point, and calculates the output when the current sampling point is input , as shown in formula (4), in the above formula, W z 、U z 、W r 、U r 、W h and U h represent

[0036] (4)

[0037] The weights of the model training, , , represents the bias, 、 Represent the sigmoid and tanh activation functions respectively, represents dot product calculation, and the time series features output by GRU are noise-free data.

[0038] S112: Slide and slice the noise-free data and set the window size m and sliding step length n, use the set step size to slide and intercept the data of the specified window length, and use the data intercepted by each slide as a slice sample. All slice samples are randomly divided into training set, validation set and test set in proportion as the model training set.

[0039] Based on the above discussion, in an optional embodiment of the present application, the window size is set m = 800, sliding step length n = 6, and the ratio of the training set, validation set, and test set is 2:2:6.

[0040] S12: Extract 4 features from slice denoising samples.

[0041] It should be noted that only extracting a single feature cannot capture the MOS tube performance type contained in the data. Considering the integrity and distinguishability of information provided by the multi-feature fusion method, this application extracts multi-features of time domain, frequency domain, cyclic bispectrum and voltage change curve parameter combination to provide richer and more comprehensive fusion information, specifically including four features: original waveform timing, maximum spectrum and corresponding frequency combination, cyclic bispectrum graphic features and power-on and power-off curve morphology description parameter combination.

[0042] S121: When there are problems with the performance of the MOS tube, the sampled drain voltage and the original timing data of the sampled voltage will have four states of "envelope sawtooth, ripple, harmonic inclusion, and normal" under certain circumstances. The denoised original slice sample data is input into the Siamese network as a feature to identify the four states, so this application selects the original slice sample as the first feature.

[0043] S122: When a MOS tube has performance problems, its sampled voltage data will be mixed with various harmonics. Different harmonic spectrum values ​​correspond to different circuit states. To extract the frequency domain state information of the sampled data, the sliced ​​samples are transformed by FFT Fourier transform to obtain the spectrum sequence. The 10 frequency points with the largest spectrum values ​​are found from the spectrum sequence values. At the same time, it is considered that the spectrum values ​​with very small values ​​have no practical significance. Only when the spectrum value reaches a certain threshold can it represent the specific frequency domain characteristics of the circuit. Through the big data statistical learning method, the abnormal thresholds of different frequency spectrum values ​​are extracted. Value max + 6δ, where Value max Represents the maximum value of the normal fluctuation range. 6δ represents 6 times the average variance of the spectrum value. Finally, a combination matrix of the form of "frequency-spectrum value-threshold" is formed as in formula (5). With this matrix as the time series input, the single feature state prediction Siamese network model gives "low-frequency harmonics,

[0044] (5)

[0045] There are four frequency domain states: "high frequency harmonics, high and low frequency harmonics, and no harmonics". This application uses the "frequency-spectrum value-threshold" combination matrix as the second feature.

[0046] S123: The sampling curves of the MOS tube drain voltage and the load current sampling voltage when the circuit is powered on and off can reflect the five states of the circuit: "delayed start, too fast start, delayed shutdown, too fast shutdown, and normal". In order to extract the sampling curve state, a set of parameter combinations that can describe the curve shape is proposed: (1) Start time t 1 , represents the time required for the sampled voltage to reach 2% of the stable voltage when power is on or off, ⑵ rise and fall time t 2 , represents the time required for the rising or falling voltage to reach 90% of the stable voltage, ⑶ Peak time t 3 , represents the time required to reach the final stable voltage value at the end of power-on and power-off, (4) delay time t 4 , represents the time required for the voltage to reach 50% of the stable voltage. [t 1, t 2, t 3, t 4 ] As the third feature.

[0047] S124: The data collected by this application belongs to a stationary random signal and contains Gaussian noise, while the circuit state belongs to a non-stationary, nonlinear, and non-Gaussian feature. In order to suppress Gaussian noise and extract the non-stationary periodic state characteristics, this application selects the second-order cyclic spectrum in the high-order statistics, and uses the unique advantage of the cyclic double spectrum that can separate the carrier and modulation information to extract the modulation frequency information of the MOS tube sampling data. Based on the cyclic double spectrum, the Siamese network identifies the four states of "amplitude modulation, phase modulation, aliasing modulation, and no modulation". In order to reduce the calculation amount of the high-order spectrum, the cyclic frequency α at the maximum peak point is selected for slicing to obtain a two-dimensional slice spectrum containing only a single frequency. Assuming x(t) is the denoised slice sample, the sampled sample of this application is a stationary voltage time series signal with a zero mean, where t=0, 1, ..., m-1, then the algorithm flow for calculating the cyclic double spectrum slice of x(t) is as follows: ① Calculate the third transformation of x(t) according to formula (6), where: i =1, 2, …, n1 , τ2 is 1, 2,…, n2 Select one of the constants, τ 1 =1, 2, …, n2 , is the lag time; m is the number of sampling points of the signal, m = n1 + n2 ; ② Calculate the third-order cyclic cumulant of the signal x(t) based on formula (7), where α is the cyclic frequency at the peak point; ③ Calculate the cyclic bispectral slice of the signal x(t) at the α frequency point based on formula (8). The cyclic bispectral slice is

[0048] (6)

[0049] (7)

[0050] (8)

[0051] The Fourier transform of the third-order cyclic cumulant at the cyclic frequency α is obtained by obtaining slice information at different frequencies, thereby reducing the amount of calculation. The slice graphic feature is the fourth feature of the present application.

[0052] S13: Establishing an EMA-Siamese single-feature state prediction model to predict circuit states based on the four single features respectively.

[0053] It should be noted that the four types of single features extracted in this application are all small samples, and it is impossible to train the model based on data-driven methods. In order to improve the circuit state prediction capability of the small sample training model, this application adopts the Siamese twin model. The Siamese network is mainly used in the field of one-shot recognition, and the vectors are classified by similarity functions. In order to further improve the feature extraction capability of the model, this application inserts a multi-scale effective attention mechanism EMA module with cross-space learning capability into the Siamese CNN infrastructure. The proposed model has the following specific implementation steps.

[0054] S131: Sample preparation, S12 extracted 4 features, which are represented as [l1, l2, l3, l4] Each feature can identify multiple circuit states. Taking the original sampling slice sample feature as an example, four states of "envelope sawtooth, ripple, harmonic inclusion, and normal" can be identified, which are represented as [s1, s2, s3, s4] . Random Selection [s1, s2, s3, s4]A sample from one of the states is taken as A (arrow sample), and another sample P (positive sample) is selected from the same state as a similar learning sample. Then, a state N (negative sample) is randomly selected from the remaining 3 states as a negative comparison sample to form a training sample set of the "A, P, N" type. The 3 slice samples selected based on the above method are respectively input into 3 identical Siamese twin networks for training.

[0055] S132: Model architecture. The EMA attention mechanism module reshapes the dimensions of some convolution channels of CNN to avoid the dimensionality reduction defects of convolution operations. At the same time, the module is designed as a multi-scale parallel sub-network architecture, and each parallel sub-network can be locally fused across channels, thereby achieving cross-network cross-fusion learning capabilities and improving the feature extraction capabilities of the CNN convolution process. Through this design pattern, the EMA model can be inserted into other neural network structures to enhance their feature mining capabilities.

[0056] S133: Network training, the CNN-EMA attention convolution channel of the Siamese twin network is similar to a feature extractor. By encoding and compressing the features of the input network, more state-representative features are further extracted. Then the distance calculation module calculates the distance between the "A, P, N" features according to the compressed encoding and Loss value, where the distance formula is d(A,P) = (A - P) 2 , d(A,N) = (A - N) 2 ,in, d The smaller the value, the closer the distance between the two features is, and the higher the probability that they belong to the same type of state. Substitute the distance calculated by the above formula into the loss function: Loss = max((d(A,P) - d(A,N) + m),0) , where the constant m is the distance difference value, m The larger the value, the when d(A,P) - d(A,N) + m < 0 is satisfied The better the model training, the better the gradient descent algorithm Loss The value is back-propagated to continuously update the network weights, and finally the state prediction Siamese small sample model is obtained.

[0057] S134: State prediction, after training, from [s1, s2, s3, s4]Randomly select two slice samples of the same state from the four states to calculate the distance. The average value obtained after multiple selections and calculations is the distance between samples of a certain state. Similarly, select two slice samples of different state types to calculate the distance. The average value is obtained after multiple calculations, and a total of 6 mean distances between different types are obtained. During testing, the slice samples are input into the Siamese model to obtain two output codes. The codes are input into the distance function to calculate the distance and then compared with the mean distance. When the distance between the two samples is less than or equal to the mean distance of the same state, they are judged to be in the same state. Otherwise, when the distance between the two samples is greater than or equal to the mean distance of different states, they are judged to be samples of different states, thereby realizing the prediction of single-feature states.

[0058] Based on the above discussion, in an optional embodiment of the present application, for the EMA-Siamese model established above, the three parallel sub-networks respectively use 1*3, 1*5 and 1*7 convolution kernels to extract feature information of different scales, and all nodes of the model use relu as the activation function. The number of "A, P, N" triplet sample sets is 200, the ratio of training samples to test samples is 5:1, the decision boundary m is set to 12000 according to the actual model operation results, the gradient descent algorithm uses the triple center group loss function to optimize the model parameters, the input data batch of each training is selected as 4, the learning rate is set to 0.005, and the number of iterations Epoch is set to 1000.

[0059] S14: The Siamese model predicts four different states of the circuit according to the four features extracted in S13, and then combines the four predicted states, uses a genetic algorithm to perform crossover and mutation operations on the state combination, and finally obtains a classification model that can evaluate the performance of the MOS tube.

[0060] It should be noted that the genetic algorithm GA generates multiple optimal classification patterns through three steps of selection, crossover and mutation, which has a good effect on global optimization tasks. This application combines the four single feature states predicted by the Siamese model into a genetic algorithm genome, and changes the gene encoding through continuous crossover and mutation, thereby mining the optimal classification pattern library.

[0061] The 4 features extracted by S12 correspond to [4,4,5,4] single feature states predicted by the S13 method. Combining the 4 states predicted each time can form 4*4*5*4, a total of 320 different state combinations. In order to mine the state combination patterns corresponding to different performance types of MOS tubes, this application establishes the optimization goal of minimizing the false alarm, false alarm and misclassification rate, taking each state type as a fragment of the genome, and using a genetic algorithm to mine the 4 state combination patterns corresponding to different MOS tube performances, which specifically includes 3 steps.

[0062] S141: Encode the gene fragment of the GA algorithm. The four positions of the gene fragment represent the Siamese model prediction states corresponding to the four features of the original data slice, the 10 highest spectrum values, the power-on and power-off parameter combination, and the cyclic bispectral slice diagram, and the value of each position represents the state type. For example: the first position is "0" represents the "envelope sawtooth" state corresponding to feature 1, the first position is "1" represents the "ripple" state, and the second position is "0" represents the "delayed start" state corresponding to feature 2. The other positions of the gene fragment are defined according to the above method.

[0063] S142: Initialize the genome population, and generate a new population after the initial population crosses and mutates. Crossover means that the same position state of different genomes is exchanged, such as Figure 3 In the figure, the spectrum values ​​and power-on / off characteristics of [1, 2, 1, 3] and [2, 3, 3, 1] cross to form the offspring [1, 3, 3, 1] and [2, 2, 1, 1]. The crossover can occur at one position or multiple positions at the same time. "Mutation" means that a certain state of a gene combination individual mutates. For example, in Figure 3, the state of the corresponding position of the highest spectrum value feature of the genome [1, 3,3, 1] mutates from "3" to "1" to form the offspring [1, 1, 3, 1] genome. Similarly, the state of the corresponding position of the power-on / off feature of the genome [2, 2, 1, 1] mutates from "1" to "3" to form the offspring [2, 2, 3, 1] genome.

[0064] S143: In order to select the best genome individual, the present application improves the genetic algorithm into four independent subspaces with parallel crossover mutation, corresponding to the performance types of four MOS tubes respectively. Each genome of the formed new offspring population is subtracted from the corresponding position of the corresponding MOS tube performance label state combination one by one, as shown in formula (9), where [g i 1 , g i 2 , g i 3 ,g i 4 ] 、 [t i 1 , t i 2 , t i 3 , t i 4 ] represent the gene fragments and

[0065] (9)

[0066] MOS tube performance label state combination, in the result of formula (9), the value of "0" represents that the gene fragment at this position is consistent with the MOS tube performance label state, and "8" represents inconsistency, where k is the number of offspring populations.

[0067] S144: This application adopts three-layer loop logic to mine the optimal state combination. ① The outermost loop is the performance type. Each loop selects one of the four performance types "excellent, good, qualified and poor". ② The second loop selects the genome individuals within the offspring population. Considering the parallel operation mechanism of the four independent subspaces proposed in this application, this loop selects a offspring population corresponding to the MOS tube performance type. ③ The innermost loop uses the selected genome individuals and all historical record samples of the outermost performance category to make a difference based on formula (9), and calculates the number of equal differences for each genome individual. n m , select each MOS tube performance type n m The top 50% of the genomes with the largest values ​​are selected as the preferred genomes.

[0068] S145: Calculate the cumulative sum of the missed alarm rate, false alarm rate, and classification error rate when the preferred genome is used as the evaluation mode R , the genetic algorithm selects the loss function value in the population R The largest 50% of gene combinations are optimized, and 50% of new state combinations are randomly generated and added to the genome population. The values ​​are recalculated through crossover and mutation operations. When the value is less than the exit threshold, the optimization ends and the optimal performance evaluation state combination pattern library is mined.

[0069] Based on the above discussion, in an optional embodiment of the present application, the initial population of each subspace of the genetic algorithm selects 200 gene combinations, the 1st, 2nd, and 4th characteristic states are represented as "1, 2, 3, 4" in the genetic algorithm gene fragment, and the 3rd characteristic state is represented as "1, 2, 3, 4, 5", and the MOS tube performance is set to 4 results of "excellent, good, qualified, and poor".

[0070] S15: Match the multi-feature state combination with the MOS tube performance evaluation mode, and identify the MOS tube performance type according to the matching result.

[0071] It should be noted that after obtaining the MOS tube performance evaluation mode library, the voltage signal of the key position of the DC load circuit is collected in real time, and the MOS tube performance can be evaluated through the above steps S11 to S14, which mainly includes 4 steps.

[0072] S151: The sampling voltage and the high-frequency digital signal collected from the drain of the MOS tube are subjected to denoising and outlier removal processing through the GRU timing network, and a training data set is obtained after sliding slicing.

[0073] S152: Obtain four features from the training data set: original slice samples, pairs of 10 highest FFT spectrum values ​​and frequencies, power-on and power-off morphology description parameter combinations, and second-order cyclic spectrum slice diagrams.

[0074] S153: Input the above features into the trained Siamese small sample model, predict the circuit single feature state corresponding to each feature, and form a state combination.

[0075] S154: matching the state combination with the MOS tube performance classification evaluation combination pattern library mined by the genetic algorithm, and performing online performance classification evaluation on the MOS tube based on the matching result.

[0076] In summary, this application proposes a method for evaluating the performance of MOS tubes for electronic loads based on multi-feature pattern recognition. The method includes high-frequency acquisition of the sampled voltage of the load circuit and the drain voltage of each MOS tube, pre-processing the sampled signal and then sliding and slicing to form a sample set; extracting 4 time-frequency domain features from the sliced ​​samples, establishing a Siamese small sample model, and predicting the single feature state of the circuit; using the single feature state as a gene fragment, using a genetic algorithm to mine the classification pattern of the performance evaluation of MOS tubes for electronic loads. This application provides a method for evaluating the performance of MOS tubes for electronic loads, screening unqualified MOS tubes, and ensuring the health and stability of the electronic load load circuit.

[0077] The present application uses specific examples to illustrate the principles and implementation methods of the present invention, and the description of the above embodiments is only used to help understand the method and core ideas of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for evaluating the performance of MOS tubes for electronic loads based on multi-feature pattern recognition, characterized in that: include: The load current sampling voltage and the drain voltage of each MOS tube are collected at high frequency, and the sampling timing signal is pre-processed and denoised, and then the sliding slice is formed into a slice sample set; Extract four features from the slice sample set: original waveform, frequency domain, power-on / off waveform morphology, and cyclic bispectral graph; The four features correspond to different single-feature circuit states, ① the original waveform corresponds to the four states of "envelope sawtooth, ripple, harmonic inclusion, and normal", ② the frequency domain features correspond to the four states of "low-frequency harmonics, high-frequency harmonics, high- and low-frequency harmonics, and no harmonics", ③ the power-on and power-off waveforms correspond to the five states of "delayed start, too fast start, delayed shutdown, too fast shutdown, and normal", and ④ the cyclic bispectrum corresponds to the four states of "amplitude modulation, phase modulation, aliasing modulation, and no modulation". The circuit state samples corresponding to the four features are divided into a ternary training set in the form of "arrow sample A, positive sample P, and negative sample N". The attention mechanism Siamese single-feature circuit state prediction model is optimized with the ternary center loss as the objective function, and the circuit state is predicted based on the four single features respectively; The predicted circuit single characteristic state is used as the gene fragment, the MOS tube classification performance is used as the label, and the genetic algorithm is used to mine the MOS tube performance classification evaluation mode for electronic loads; The performance classification and evaluation of MOS tubes in DC load circuits are carried out based on the evaluation model mined by genetic algorithm.

2. The method for evaluating the performance of a MOS tube for an electronic load based on multi-feature pattern recognition according to claim 1, characterized in that: The load current sampling voltage and the drain voltage of each MOS tube are collected at high frequency, and the sampling timing signal is pre-processed and denoised, and then the sliding slice is formed into a slice sample set, including: Based on the least squares objective function, a multi-input and multi-output GRU timing network is trained to remove the random noise generated by sampling, and a sliding slicing technique is used to generate a slice sample set for the high-frequency sampling data after the random noise is removed.

3. The method for evaluating the performance of a MOS tube for an electronic load based on multi-feature pattern recognition according to claim 1, characterized in that: Four features are extracted from the slice sample set: original waveform, frequency domain, power-on / off waveform morphology, and cyclic bispectral graph, including: Four features of the original collected data are extracted: ① the denoised slice samples are taken as the first feature, ② the slice samples are subjected to FFT transformation, and the "frequency-spectrum value-threshold" pairs consisting of the 10 highest spectrum values ​​and the corresponding frequencies are selected as the second feature, ③ the combination of parameters describing the power-on and power-off sample curves is extracted as the third feature, and ④ the cyclic bispectral slice diagram of the slice samples is drawn as the fourth feature.

4. The method for evaluating the performance of a MOS tube for an electronic load based on multi-feature pattern recognition according to claim 1, characterized in that: The attention mechanism Siamese single-feature circuit state prediction model is optimized with the ternary center loss as the objective function, and the circuit state is predicted based on the four single features, including: A one-shot feature matching algorithm is selected, and a ternary sample set in the form of "state A, similar state P with different working conditions, and different state N" is established. The ternary center loss function is selected to train the attention mechanism Siamese single-feature circuit state prediction model. The extracted four features are used as input, and a 1D or 2D convolution method is selected according to the input sample form. The model predicts the single-feature state of the corresponding input. The attention mechanism Siamese single-feature circuit state prediction model inserts a multi-scale effective attention mechanism EMA module with cross-space learning ability into the Siamese model based on CNN to enhance the feature extraction ability of the network.

5. The method for evaluating the performance of a MOS tube for an electronic load based on multi-feature pattern recognition according to claim 1, characterized in that: The predicted circuit single characteristic state is used as the gene fragment, the MOS tube classification performance is used as the label, and the genetic algorithm is used to mine the MOS tube performance evaluation classification mode for electronic loads, including: The single feature state is used as a segment of the 4-segment genome of the genetic algorithm. The location and content of the gene segment are numerically encoded according to the MOS tube performance and the single feature circuit state. The circuit state combination pattern library for classifying MOS tube performance is selected through crossover mutation.

6. The method for evaluating the performance of a MOS tube for an electronic load based on multi-feature pattern recognition according to claim 1, characterized in that: The evaluation model based on genetic algorithm mining is used to perform performance classification evaluation on MOS tubes in DC load circuits, including: extracting four features from denoised slice samples collected online as input of the Siamese single-feature circuit state prediction model of the attention mechanism, predicting a circuit single-feature state for each feature input through the model's ternary feature matching method, combining the four single-feature states and matching them with the MOS tube performance evaluation pattern library, and providing a performance classification evaluation of the MOS tube based on the correspondence between the matching pattern and the MOS tube performance type.

Images