Gallium nitride E-type inverter circuit design and optimization method based on ZVS feature extraction

Abstract

The invention discloses a method for designing and optimizing a gallium nitride E-type inverter circuit based on ZVS (Zero Voltage Suppressor) feature extraction, aiming at the limitation that the dynamic characteristics of the circuit are difficult to accurately capture in the traditional method, the ZVS features are innovatively extracted from a sample generated by Monte Carlo simulation, and the subtle change of ZVS waveforms is comprehensively reflected. The ZVS features and device parameters are integrated, an enhanced data set is constructed, a neural network model is used for training, and high-precision mapping among the ZVS features, output power, conversion efficiency and circuit parameters is achieved. The introduction of the ZVS features overcomes the defect that a traditional method depends on manual extraction of ZVS information, rapid optimization of circuit parameters is realized, the design period is greatly shortened, and the reliability and the overall performance of the circuit are improved. Experimental results show that compared with a traditional method, the method is remarkably improved in the aspects of prediction precision and conversion efficiency optimization.

Description

Technical Field

[0001] The present invention belongs to the technical field of power switching devices in power electronics technology, and in particular relates to a design and optimization method of a gallium nitride E-class inverter circuit based on ZVS feature extraction. Background Art

[0002] With the increasing prominence of energy crisis and environmental problems, efficient power conversion technology has become increasingly important. Power electronics technology aims to achieve efficient conversion, control and application of electric energy, and inverter is one of the important circuits. Due to the breakthrough progress in the performance of new semiconductor power devices, Class E inverter, as an efficient power conversion solution, has attracted widespread attention from academia and industry due to its high efficiency and simple structure. The inverter converts DC power into high-frequency AC power to drive the transmitting coil to generate an electromagnetic field, while the Class E inverter minimizes switching losses through ZVS technology, which is particularly suitable for high-frequency applications and can effectively drive the transmitting coil. Improving the conversion efficiency of Class E inverter is crucial, as it can reduce energy loss, improve transmission efficiency, and reduce operating costs to meet various needs. However, the key performance indicators of traditional Class E amplifiers, such as efficiency, output power and stability, will fluctuate significantly under slight deviations in component parameters and loads. At the same time, problems such as high stress operation of the device, limited bandwidth and parasitic effects make traditional Class E amplifiers face many challenges in practical engineering applications. Therefore, seeking effective improvement solutions to improve the robustness and applicability of Class E amplifiers is a key issue that needs to be solved urgently.

[0003] With the urgent demand for high efficiency, high power density and high frequency in applications such as wireless power transmission and RF power amplifiers, traditional silicon-based semiconductor devices have gradually shown performance bottlenecks. Third-generation semiconductor materials such as gallium nitride (GaN) have shown great potential in the design of class E inverter circuits due to their unique physical properties, higher electron mobility, lower on-resistance and faster switching speed. Compared with traditional silicon devices, GaN devices can significantly increase the operating frequency of class E inverter circuits, thereby reducing the volume and weight of the circuit, and effectively improve the conversion efficiency and reduce energy loss. However, high-frequency operation also brings a series of challenges, the most important of which is the significant enhancement of parameter sensitivity. At high frequencies, slight deviations in component parameters and loads will lead to significant decreases in efficiency, output power fluctuations and even circuit imbalance. Therefore, how to optimize the design and effectively control the GaN class E inverter circuit and overcome the design difficulties caused by parameter sensitivity is an important direction of current research.

[0004] In the field of circuit hardware analysis, the Monte Carlo method, as an important analysis tool, is widely adopted to improve the yield and efficiency of circuit design. The present invention utilizes the Monte Carlo analysis method. By randomly sampling the device parameters in the circuit, a large number of different parameter combinations are simulated, and circuit simulations are carried out, so as to obtain the characteristic data of the circuit under different parameter combinations, such as output power, efficiency, etc., and particularly focus on the data in the zero-voltage switching state. ZVS, as a key operating state in circuits such as switching power supplies, directly affects the conversion efficiency and reliability of the circuit. In traditional methods, the extraction of the ZVS state usually relies on manual waveform observation or complex post-processing, which is inefficient and error-prone, and it is difficult to perform large-scale automated analysis. In addition, the ZVS state itself is also affected by various complex factors, such as small deviations in device parameters, circuit topology, operating frequency, etc., which makes it very difficult to accurately predict the ZVS state and its impact on the conversion efficiency. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide a design and optimization method for a gallium nitride class-E inverter circuit based on ZVS feature extraction. It overcomes the disadvantages of traditional methods that rely on manual extraction of ZVS information and insufficient prediction accuracy, and provides a more effective means for circuit optimization design.

[0006] Technical Solution: A design and optimization method for a gallium nitride class-E inverter circuit based on ZVS feature extraction according to the present invention is characterized by including the following steps:

[0007] S1: Determine the initial device parameter values included in the gallium nitride-based class-E inverter circuit;

[0008] S2: Set the initial tolerance of a single device and perform multiple Monte Carlo simulations to obtain a number of samples generated by the Monte Carlo simulation;

[0009] S3: Extract the ZVS data features for each sample generated by the Monte Carlo simulation, and combine the feature data with the initial device parameter values to form a data set;

[0010] S4: Based on the data set, create neural network models MLPⅠ and MLPⅡ adapted to the circuit;

[0011] S5: Select highly important features from the device parameters obtained from the Monte Carlo simulation and the extracted ZVS sample features, and use feature crossing to increase the feature input items of the neural network and expand the feature space of the model;

[0012] S6: Standardize the crossed training set as the input of the neural network models MLPⅠ and MLPⅡ, and keep the power and conversion efficiency obtained from the Monte Carlo simulation unchanged as the values to be predicted;

[0013] S7: Evaluate the model performance and modify the circuit component parameters according to the device parameters corresponding to the optimal conversion efficiency output by the model.

[0014] Furthermore, step S1 is specifically as follows: The circuit structure of the gallium nitride class-E inverter circuit based on ZVS feature extraction consists of a choke inductor L1, a bypass capacitor C1, resonant elements L2 and C2, a GaN power switch device S, L3 and C3 that form a parallel filter circuit, and a resistor R.

[0015] The formula for the DC input voltage is as follows:

[0016]

[0017] The input power P in is calculated by the following formula:

[0018] P in = V in I

[0019] After determining the resistance R of the circuit, the output power of the circuit is:

[0020]

[0021] where

[0022] The conversion efficiency is thus obtained as:

[0023]

[0024] The following formula for calculating the parameter of capacitor C1 in the class-E inverter circuit is obtained:

[0025]

[0026] Furthermore, the expression for the resonant capacitor C2 is obtained as:

[0027]

[0028] where Q L is the load quality factor;

[0029] The parameter value of the inductor L1 is:

[0030]

[0031] In addition, in the additionally added parallel filter, the calculation formulas for the inductor L3 and the capacitor C3 are as follows:

[0032]

[0033] For each component L1, C1, L2, C2, L3, and C3 in the circuit, a tolerance range of 5 - 20% is set, and Monte Carlo simulation analysis is performed.

[0034] Further, step S2 is specifically as follows: Randomly sample the parameters of each component in the class-E inverter circuit by the Monte Carlo method, and combine the conversion efficiency data obtained from the simulation to determine the optimal value range of each device parameter. After determining the optimal range of each device parameter, simulate the entire circuit by the Monte Carlo method to obtain the corresponding relationship data between several device parameters and circuit performance.

[0035] Further, in step S3, the ZVS data feature extraction includes GS and DS various statistical indicators such as the maximum value, mean value, kurtosis, and skewness of the curve.

[0036] Further, step S4 is specifically as follows: Train a neural network using the dataset combined by Monte Carlo simulation and ZVS feature data. Build a multi-layer perceptron through the Pytorch framework. The topological structure of the neural network model MLPⅠ includes an input layer with six input channels, corresponding to six input features C1, C2, C3, L1, L2, L3, five hidden layers each containing 64, 32, 32, 16, and 8 neurons, respectively, and an output layer; the input layer and hidden layers use the ELU activation function, and the output layer uses the linear activation function;

[0037] After the neural network model MLPⅠ is built, build the neural network model MLPⅡ using the Pytorch framework. Its topological structure includes an input layer with eighteen input channels, corresponding to eighteen input features, including the features C1, C2, C3, L1, L2, L3 of the circuit components and the state features V ds,max , ds,min , ds,mean , ds,devs , ds,kurtosis , ds,skewness , gs,max , gs,min , gs,max , gs,mean , gs,kurtosis , gs,skewness extracted by ZVS, seven hidden layers each containing 128, 64, 32, 32, 16, and 8 neurons, respectively, and an output layer. Among them, the input layer and hidden layers use the ELU activation function, and the output layer uses the Linear linear activation function.

[0038] Use a multi-output random forest regression model to predict two outputs simultaneously, calculate the average importance of features for these two outputs of output power and conversion efficiency, rank the features in descending order of importance, and select the 4 features with greater importance, namely V ds,skewness , V ds,mean , V ds,kurtosis , V gs,max .

[0039] Furthermore, step S5 is specifically as follows: Use a multi-output random forest regression model to predict two outputs simultaneously, calculate the average importance of features for these two outputs of output power and conversion efficiency, perform importance ranking on circuit parameter features and ZVS features simultaneously, select the features with greater importance for feature crossing, increase the feature input items of the neural network, and expand the feature space of the model.

[0040] Furthermore, step S6 is specifically as follows: Standardize the sample features of the enhanced dataset, selectively use them as the inputs of neural networks MLPⅠ and MLPⅡ, and train the neural network model.

[0041] Furthermore, step S7 is specifically as follows: Use the early stopping method to retain the model with the best training, adopt MES and percentage error to evaluate the accuracy of the model, and accurately adjust the circuit parameters according to the maximum conversion efficiency predicted by the model to optimize the circuit performance.

[0042] As Figure 4 , ELU is a smooth and continuous function, easy to calculate the gradient, and it is necessary to standardize the input data using methods such as Z-score normalization at the input layer to improve the learning efficiency and training stability of the model.

[0043] In the above design scheme, for Z-score normalization, the performance evaluation of the model uses mean squared error (MSE) and percentage error, which are calculated as follows:

[0044]

[0045] Among them, x is the original input feature, μ is the average value of the input feature, σ is the standard deviation of the input feature, and z is the input feature after standardization.

[0046]

[0047] Among them, n represents the total number of prediction samples, y i represents the true value of the i-th sample, represents the predicted value of the i-th sample. The smaller the value of MSE, the closer the predicted result of the model is to the true value, and the higher the prediction accuracy. MAPE is an index used to measure the prediction of the regression model. The smaller this index, the higher the prediction accuracy of the model.

[0048] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages:

[0049] (1) Through feature crossing, the present invention expands the model feature space, mines high-order features containing ZVS information, enhances the non-linear expression ability of the model, makes more full use of the prediction ability of ZVS features for conversion efficiency, further improves the model prediction performance and generalization ability, and finally can optimize the circuit design according to the optimal parameters and ZVS state output by the model to achieve higher conversion efficiency.

[0050] (2) The ZVS feature significantly improves the prediction accuracy of conversion efficiency: The present invention innovatively extracts features directly reflecting the ZVS state and integrates them with device parameters into a new data set, and uses MLPⅠ to train this data set to achieve accurate prediction of ZVS state data; at the same time, MLPⅡ is used to train the same data set to predict the conversion efficiency of the circuit. By introducing the key index of ZVS, the present invention significantly improves the prediction accuracy of the neural network model for conversion efficiency, enables the model to more accurately capture the influence of ZVS on efficiency, and provides a more reliable basis for circuit optimization.

[0051] (3) The advantage of the present invention using neural network lies in its powerful non-linear fitting ability, which can effectively learn high-dimensional and non-linear circuit characteristic data generated by Monte Carlo simulation and accurately predict the circuit performance under different parameter combinations. The characteristics of the present invention using neural network are: first, there is no need to establish a complex circuit analysis model, and only need to learn through a data-driven method, which simplifies the modeling process; second, it can quickly predict the circuit performance under a large number of parameter combinations, greatly improving the optimization efficiency and shortening the design cycle; third, it can more accurately capture the complex non-linear relationship between device parameters, ZVS features and conversion efficiency, overcomes the limitations of traditional methods in dealing with non-linear problems, and thus more effectively optimizes the circuit design. Description of the drawings

[0052] Figure 1 It is a schematic diagram of a design and optimization method for a gallium nitride class-E inverter circuit based on ZVS feature extraction according to the present invention.

[0053] Figure 2 It is a schematic diagram of the structure of a common-source common-gate gallium nitride class-E inverter circuit.

[0054] Figure 3 It is a schematic diagram of a multi-layer perceptron network structure.

[0055] Figure 4 It is a discount comparison diagram of the true value - predicted value of conversion efficiency.

[0056] Figure 5 It is to output the ZVS feature V DSComparison chart of the actual value of kurtosis and the predicted value line

[0057] Figure 6 Output ZVS characteristic V GS Comparison chart of the actual value of kurtosis and the predicted value line.

[0058] Figure 7 Comparison chart of the output power prediction and true value of the model with and without ZVS feature training.

[0059] Figure 8 Comparison chart of conversion efficiency prediction and true value when the model is trained with and without ZVS features. DETAILED DESCRIPTION

[0060] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0061] The present invention proposes a novel method for optimizing the circuit parameters of a gallium nitride class E inverter circuit. First, a large number of initial circuit device parameter combinations are randomly generated within a preset parameter range through Monte Carlo simulation, and circuit simulation is performed to obtain corresponding conversion efficiency data and build an initial database. The key innovation is to extract characteristic data that directly reflects the ZVS state from the simulation results, such as the V GS and V DS The waveform's maximum, mean, standard deviation, kurtosis and skewness are extracted, and the feature space is expanded through feature crossover to mine high-order features containing ZVS information. Then, two neural network models are used: Model I takes the integrated data set (including device parameters and ZVS state data) as input to train and predict ZVS state data; Model II also takes the integrated data set as input to train and predict conversion efficiency, and the input of Model II includes the ZVS state data predicted by Model I, so as to more accurately capture the nonlinear effect of ZVS on conversion efficiency. The introduction of neural networks, in particular, highlights its powerful nonlinear fitting ability, which can effectively capture the complex nonlinear relationship between circuit parameters, ZVS state and conversion efficiency, process high-dimensional data, and show good generalization ability. Finally, according to the conversion efficiency predicted by Model II and the ZVS state predicted by Model I, the optimal device parameter combination is selected, so that the circuit design can be optimized efficiently and accurately without a large number of repeated simulations, and the conversion efficiency can be significantly improved.

[0062] A design and optimization method for a gallium nitride class E inverter circuit based on ZVS feature extraction includes the following principles:

[0063] The circuit structure of the GaN class E inverter circuit based on ZVS feature extraction is as follows: Figure 2As shown in the figure, it consists of a choke inductor L1, a bypass capacitor C1, resonant elements L2 and C2, a GaN power switch device S, L3 and C3 that form a parallel filter circuit, and a resistor R.

[0064] The formula for the DC input voltage is as follows:

[0065]

[0066] Input power P in can be calculated by the following formula:

[0067] P in = V in I# (5)

[0068] After determining the resistance R of the circuit, the output power of the circuit is obtained as:

[0069]

[0070] where

[0071] From this, the conversion efficiency can be obtained as:

[0072]

[0073] Through formula (6) and the resistance R, the following formula for calculating the parameter of capacitor C1 in the class-E inverter circuit can be obtained:

[0074]

[0075] From formula (8) and formula (9), the expression for the resonant capacitor C2 is obtained as:

[0076]

[0077] where Q L is the load quality factor.

[0078] The parameter value of the inductor L1 is:

[0079]

[0080] In addition, in the additionally added parallel filter, the calculation formulas for the inductor L3 and the capacitor C3 are as follows:

[0081]

[0082] For each component L1, C1, L2, C2, L3, and C3 in the circuit, a tolerance range of 5 - 20% is set, and Monte Carlo simulation analysis is carried out. In each simulation, the component parameters vary randomly within their tolerance ranges, and the corresponding conversion efficiency is recorded to form a data set. By analyzing the data set, the results with higher conversion efficiency are screened out, and the corresponding component parameter values are found. Then, taking these screened parameter values as the new initial values and narrowing the tolerance range, Monte Carlo simulation is carried out again for iterative optimization. Through multiple iterations, the optimal parameter combination is gradually approached, and finally the optimal device parameter value combination corresponding to a conversion efficiency greater than 85% is found.

[0083] According to Figure 3 the structure shown, a neural network is trained using a data set merged from Monte Carlo simulation and ZVS characteristic data. A multi-layer perceptron is built through the Pytorch framework. The topological structure of the neural network model MLPⅠ includes an input layer with six input channels, corresponding to six input features C1, C2, C3, L1, L2, L3, five hidden layers with 64, 32, 32, 16, and 8 neurons respectively, and an output layer. The ELU activation function is used for the input layer and hidden layers, and the linear activation function is used for the output layer. Other hyperparameters of the model are set as follows: Adam optimizer, learning rate 0.001, batch size 120, and the data set is divided into training set, validation set, and test set in the ratio of 6:2:2.

[0084] After the neural network model MLPⅠ is built, the neural network model MLPⅡ is built using the Pytorch framework. Its topological structure includes an input layer with eighteen input channels, corresponding to eighteen input features, including the features of circuit components C1, C2, C3, L1, L2, L3 and the state features V ds,max , V ds,min , V ds,mean , V ds,devs , V ds,kurtosis , V ds,skewness , V gs,max , V gs,min , V gs,max , V gs,mean , V gs,kurtosis , V gs,skewness . Seven hidden layers with 128, 64, 32, 32, 16, and 8 neurons respectively, and an output layer. Among them, the ELU activation function is used for the input layer and hidden layers, and the Linear activation function is used for the output layer. Other hyperparameters are selected as follows: optimizer Adam, learning rate 0.001, batch size 120, training set: validation set: test set = 6:2:2.

[0085] Use a multi-output random forest regression model to predict two outputs simultaneously, and calculate the average importance of features for the two outputs of output power and conversion efficiency. Then sort the features in descending order of importance and select the 4 features with greater importance, namely V ds,skewness , V ds,mean , V ds,kurtosis , V gs,max .

[0086] Feature crossing increases the input features of the model. The total number of feature items in the original Monte Carlo simulation and the state feature items of ZVS is 18, and the newly added feature items are as follows.

[0087] Select V ds,skewness , V ds,mean , V ds,kurtosis , V gs,max and other 4 features are crossed with 18 features respectively to generate 72 new crossed features.

[0088] After adopting feature crossing, the number of input feature items of the model increases from the original 18 to 90.

[0089] Perform Z-score standardization on the parameter data in the dataset, convert the data into a normal distribution with a mean of 0 and a variance of 1, and use it as the input feature of the neural network; use the state feature of ZVS as the target to be predicted by MLPⅠ, and output power and conversion efficiency as the targets to be predicted by model MLPⅡ.

[0090] Use MSE and MAPE to measure the accuracy of the model. MSE measures the accuracy of the conversion efficiency prediction. The smaller its value, the smaller the deviation between the model prediction value and the true value, and the higher the prediction accuracy; MAPE measures the prediction error in the form of a percentage. The smaller its value also represents the higher the prediction accuracy of the model.

[0091] Implementation column

[0092] A design and optimization method for a gallium nitride class-E inverter circuit based on ZVS feature extraction in this implementation column is as follows in detail:

[0093] Step C1: Build a basic circuit in the simulation software according to the Figure 2 shown circuit schematic diagram.

[0094] Step C2: Set the input voltage V in of the circuit to 20V, the operating frequency to 1MHz, and the required output power to 80W, and then successively through formulas (6), (7), (8), (9), (10), (11), (12), (13), and

[0095] The initial value of the load resistor R is calculated to be 2.884 Ω, the initial value of the capacitor C1 is 10.15 nF, the initial value of the capacitor C2 is 9.44 nF, the initial value of the capacitor C3 is 0.39 μF, the initial value of the inductor L2 is 3.21 μH, and the initial value of the inductor L3 is 66 nH. In addition, the initial value of the inductor L1 is taken as 5 μH here.

[0096] Step C3: The initial conversion efficiency of the circuit is simulated and calculated to be 91.747%. Then, the tolerance of the capacitor C1 is set to 5%, and 10,000 Monte Carlo simulation analyses are performed. The parameters of each device and the conversion efficiency data set are simulated, and the first 2,000 groups of data with higher conversion efficiency are sorted out. Thus, the data range of the capacitor C1 is determined to be 9.6425 - 10.6575 nF.

[0097] Step C4: Then, the simulation analyses are sequentially performed on the capacitors C2, C3, L1, L2, and L3. Finally, when the circuit conversion efficiency is optimal, the parameter ranges of the devices are determined to be 8.968 - 9.912 nF, 0.3705 - 0.4095 μF, 4.75 - 5.25 μH, 3.0495 - 3.3705 μH, and 62.7 - 69.3 nH, respectively.

[0098] Step C5: The two constructed neural network models MLPⅠ and MLPⅡ are used to train the data set to establish an accurate mapping relationship between the circuit parameters, ZVS characteristics, and circuit performance. The MLPⅠ model aims to learn the association between the circuit parameters and ZVS characteristics. Its input is the circuit parameters in the data set, and the output is the corresponding ZVS characteristic index. The MLPⅡ model aims to learn the relationship between the circuit parameters, ZVS characteristics, output power, and conversion efficiency. Its input is the combination of the circuit parameters and ZVS characteristic indexes in the data set, and the output is the corresponding output power and conversion efficiency.

[0099] Step C6: As the amount of training data increases, the MSE of the training set and the test set usually gradually decreases and finally stabilizes, indicating that the model gradually fits the data better and the generalization ability is enhanced. When the amount of training data reaches 4,000, the MSE of the training set is about 5.49×10 -6 or so, and the MSE of the test set is about 5.76×10 -6 or so, and the percentage error is 0.236%, indicating that the error is extremely small and the model accuracy is high.

[0100] Step C7: Check the prediction results of the MLP model. Figure 5, 6 is the comparison graph of the true value - predicted value discount of the output ZVS feature. To analyze the effectiveness of the model prediction, 100 samples were continuously extracted from a total of 1000 prediction data, and these two graphs were plotted after arranging them in ascending order according to the magnitude of the predicted values. Although the corresponding true values do not strictly increase due to their own laws and show certain local fluctuations, their overall change trend should be consistent with the predicted values, that is, when the predicted value increases, the true value also tends to increase; vice versa. The absolute error curve also tends to be stable and the values are relatively low. This indicates that the prediction effect of the model on these samples is good and it has a high prediction accuracy. Figure 7 , 8 is the comparison graph of the predicted value and the true value of the output for the model trained with and without the ZVS feature. It can be seen from the graph that the predicted value and the true value of the ZVS prediction are highly consistent, indicating that compared with the model trained without using the ZVS feature, the model trained with the ZVS feature can accurately predict the conversion efficiency. It proves that the introduction of the ZVS feature significantly improves the prediction accuracy of the model, and the model can well track the true value regardless of how the conversion efficiency fluctuates.

[0101] Step C8: The maximum conversion efficiency found according to the prediction result of the model is 96.2%, and the corresponding device parameter values are C1 = 15.44 nF, C2 = 23.62 nF, C3 = 0.17 nF, L1 = 4.86 μH, L2 = 1.55 μH, L3 = 146.05 nH. Modify the initial parameters of the circuit devices according to these parameters to make the working efficiency of the circuit reach the optimal.

[0102] The specific embodiments described above have further detailed the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A design and optimization method for a gallium nitride class-E inverter circuit based on ZVS feature extraction, characterized in that, It includes the following steps: S1: Determine the initial device parameter values included in the GaN-based Class-E inverter circuit; S2: Set the initial tolerance of a single device and perform multiple Monte Carlo simulations to obtain several samples generated by the Monte Carlo simulation; S3: Extract the ZVS data features for each sample generated by the Monte Carlo simulation, and combine the feature data with the initial device parameter values to form a data set; S4: Based on the data set, create neural network models MLPⅠ and MLPⅡ adapted to the circuit; S5: Select highly important features from the device parameters obtained from the Monte Carlo simulation and the extracted ZVS sample features, and use feature crossing to increase the feature input items of the neural network and expand the feature space of the model; S6: Standardize the crossed training set as the input of the neural network models MLPⅠ and MLPⅡ, and keep the power and conversion efficiency obtained from the Monte Carlo simulation unchanged as the values to be predicted; S7: Evaluate the model performance, and modify the circuit element parameters according to the device parameters corresponding to the optimal conversion efficiency output by the model.

2. The design and optimization method of a gallium nitride class-E inverter circuit based on ZVS feature extraction according to claim 1, characterized in that Step S1 is specifically as follows: The circuit structure of the GaN-based Class-E inverter circuit for ZVS feature extraction consists of a choke inductor L1, a bypass capacitor C1, resonant elements L2 and C2, a GaN power switch device S, L3 and C3 forming a parallel filter circuit, and a resistor R; The formula for the DC input voltage is as follows: Input power P in is calculated by the following formula: P in = V in I After determining the resistance R of the circuit, the output power of the circuit is obtained as: Among them Thus, the conversion efficiency is obtained as: The following formula for calculating the parameter of the capacitor C1 of the Class-E inverter circuit is obtained: Furthermore, the expression of the resonant capacitor C2 is obtained as: where Q L is the load quality factor; The parameter value of the inductor L1 is: In addition, in the additionally added parallel filter, the calculation formulas for the inductor L3 and the capacitor C3 are as follows: For each component L1, C1, L2, C2, L3, and C3 in the circuit, set a tolerance range of 5-20%, and perform Monte Carlo simulation analysis.

3. The design and optimization method of a gallium nitride class-E inverter circuit based on ZVS feature extraction according to claim 1, characterized in that, Step S2 is specifically as follows: Randomly sample the parameters of each component in the Class-E inverter circuit by the Monte Carlo method, and combine the conversion efficiency data obtained from the simulation to determine the optimal value range of each device parameter. After determining the optimal range of each device parameter, perform simulation on the entire circuit by the Monte Carlo method to obtain the corresponding relationship data between several device parameters and circuit performance.

4. The design and optimization method of a gallium nitride class-E inverter circuit based on ZVS feature extraction according to claim 1, characterized in that In step S3, the ZVS data feature extraction includes the statistics of the maximum value, mean value, kurtosis, and skewness of the V GS and V DS curves, and the data features are combined with the original device parameter values to form a data set.

5. The design and optimization method of a gallium nitride class-E inverter circuit based on ZVS feature extraction according to claim 1, characterized in that, Step S4 is specifically as follows: Train a neural network using the data set combined with the Monte Carlo simulation and ZVS feature data. Build a multi-layer perceptron through the Pytorch framework. The topological structure of the neural network model MLPⅠ includes an input layer with six input channels, corresponding to six input features C1, C2, C3, L1, L2, L3, five hidden layers with 64, 32, 32, 16, and 8 neurons respectively, and an output layer; the ELU activation function is used for the input layer and the hidden layers, and the linear activation function is used for the output layer; After the neural network model MLPⅠ is built, the neural network model MLPⅡ is built using the Pytorch framework. Its topological structure includes an input layer with eighteen input channels, corresponding to eighteen input features, including the features C1, C2, C3, L1, L2, L3 of circuit elements and the state features V ds,max 、V ds,min 、V ds,mean 、V ds,devs 、V ds,kurtosis 、V ds,skewness 、V gs,max 、V gs,min 、V gs,max 、V gs,mean 、V gs,kurtosis 、V gs,skewness , seven hidden layers each containing 128, 64, 32, 32, 16, and 8 neurons, and an output layer. The ELU activation function is used for the input layer and the hidden layers, and the Linear linear activation function is used for the output layer.

6. The design and optimization method of a gallium nitride class-E inverter circuit based on ZVS feature extraction according to claim 1, characterized in that, Step S5 is specifically as follows: Use a multi-output random forest regression model to simultaneously predict two outputs, calculate the average importance of features for the two outputs of output power and conversion efficiency, arrange the features in descending order of importance, and select the 4 features with greater importance, namely V ds,skewness , V ds,mean , V ds,kurtosis , V gs,max .

7. The design and optimization method of a gallium nitride class-E inverter circuit based on ZVS feature extraction according to claim 1, characterized in that, Step S6 specifically is: Standardize the cross - processed training set, standardize the device parameters obtained from Monte Carlo simulation and the extracted ZVS sample features, which are used as the inputs of neural network models MLPⅠand MLPⅡ. The power and conversion efficiency obtained from Monte Carlo simulation remain unchanged and are used as the values to be predicted.

8. The design and optimization method of a gallium nitride class-E inverter circuit based on ZVS feature extraction according to claim 1, characterized in that Step S7 specifically is: Use the early stopping method to retain the model with the best training results. Adopt MES and percentage error to evaluate the accuracy of the model. According to the maximum conversion efficiency predicted by the model, precisely adjust the circuit parameters to optimize the circuit performance.

Images