Microwave power transistor S parameter reverse modeling method based on BiLSTM-LSTM technology

Abstract

The invention discloses a reverse modeling method for an S parameter of a microwave power transistor based on a bidirectional long short-term memory neural network-long short-term memory neural network (BiLSTM-LSTM) technology, and the method comprises the five steps of data preparation, model design, hyper-parameter setting, model training and optimization, and model evaluation. According to the BiLSTM-LSTM modeling method, the core is that forward and backward and context features of data are extracted through a BiLSTM layer, time sequence dependence modeling is carried out in combination with an LSTM layer, and therefore the nonlinear mapping relation between performance parameters and S parameters of a microwave power transistor is fully excavated. According to the method, the spatial-temporal characteristics and relevance of data can be effectively captured, the precision and generalization ability of a transistor model are remarkably improved, and an efficient and accurate modeling method is provided for realizing high-performance and high-reliability circuit design.

Description

Technical Field

[0001] This invention relates to the field of microwave power transistor parameter modeling research, and more specifically, to a method for inverse modeling of microwave power transistor S-parameters based on BiLSTM-LSTM technology. Background Technology

[0002] With the rapid development of radio frequency and microwave technology towards higher reliability and higher integration, the accurate and efficient characterization of microwave power transistors is of great significance for circuit system design. Traditional microwave power transistor modeling methods (such as physical equivalent circuit models and lookup table methods) often face limitations such as model complexity, large data requirements, high cost, and sensitivity to process fluctuations when mapping high-dimensional nonlinear relationships of S-parameters across a wide frequency band in high-frequency scenarios. Therefore, Artificial Neural Networks (ANNs), with their powerful high-dimensional nonlinear fitting capabilities, provide a new technical path for transistor modeling, especially suitable for inverse extraction scenarios where traditional physical models are difficult to accurately describe.

[0003] Inverse modeling of microwave power transistors is a method that inversely derives their internal physical or operating parameters based on external performance parameters (S-parameters). Currently, ANN-based inverse modeling of microwave power transistor S-parameters generally suffers from limited data utilization efficiency and high experimental costs. Therefore, this invention employs a BiLSTM-LSTM hybrid artificial neural network method to accurately characterize the complex nonlinear relationship between microwave power transistor performance parameters and S-parameters. By monitoring changes in S-parameters, it predicts the device's operating state and can be combined with EDA software such as ADS and HFSS to automatically and rapidly optimize design parameters based on the target response, thereby effectively shortening the product development cycle. This provides a more efficient and convenient means for extracting device characteristics and inferring states in the design of RF microwave circuits. Summary of the Invention

[0004] The purpose of this invention is to provide a reverse modeling method for S-parameters of microwave power transistors based on BiLSTM-LSTM technology, so as to establish an accurate device parameter model and improve the efficiency of circuit system design.

[0005] The technical method and solution of the present invention for solving the above-mentioned microwave power transistor parameter modeling is as follows: a reverse modeling method for microwave power transistor S-parameters based on BiLSTM-LSTM technology, including data preparation, model design, hyperparameter setting, model training and optimization, and model evaluation; Step one, data preparation, involves obtaining the dataset through experimental measurements, where the input variables are the S-parameters [ReS] of GaAs pHEMT. 11 ImS 11 ReS21 ImS 21 ReS 12 ImS 12 ReS 22 ImS 22 The output variables are composed of the corresponding microwave power transistor performance characteristic parameters X = [X1, X2, X3, X4]. The dataset is then randomly divided into training and testing sets in a 4:1 ratio, and the data is normalized and preprocessed. Finally, it is transformed into a cell array suitable for training time-series networks. Step 2 involves designing the model by using the layerGraph function in the Matlab deep learning framework to create a BiLSTM-LSTM structure after determining the input and output data, and then setting the number of neurons and initial weights in the input, hidden, and output layers of the neural network. Step 3: The hyperparameters are set as follows: the maximum number of iterations is set to 30, the initial learning rate and the learning rate decrease factor are 0.01 and 0.05 respectively, and then the Adam optimizer is used to update the weights. Step four, model training and optimization, is to end the training process when the loss function value no longer decreases significantly or the number of iterations reaches the maximum value of 30. At the same time, the mean square error (MSE) is continuously monitored during model training. If the MSE value drops below 0.01, the process proceeds to step five, model evaluation; otherwise, it returns to step two, model design, adjusts the number of hidden layer neurons, and retrains. Step 5 involves model evaluation using MSE, Mean Absolute Error (MAE), and Coefficient of Determination (R-Square, R0). 2 () is used as a regression index to evaluate the model's performance.

[0006] The beneficial effects of the technical solution of this invention are as follows: This invention adopts the BiLSTM-LSTM modeling method, the core of which is that it has both bidirectional and unidirectional time series modeling capabilities, which can fully extract and utilize the contextual information in the data and significantly improve the feature representation capability; secondly, the model has a strong nonlinear fitting capability, which can effectively capture the spatiotemporal features and correlations in the data, enabling it to accurately represent the characteristic parameters of microwave power transistors under the condition of a small amount of data, which can significantly improve the accuracy and generalization ability of the transistor model, and provide an efficient and accurate modeling method for realizing high-performance and high-reliability circuit design.

[0007] Furthermore, the characteristic parameter X = [X1, X2, X3, X4] in step one, where X1 is the frequency (f) and X2 is the gate current (I). gs X3 is the drain voltage (V)ds X4 is the gate voltage (V) gs ).

[0008] The beneficial effects of the above-mentioned further solutions are: the output parameters of the present invention are the most core and complete set of physical quantities of the DC bias point and high-frequency operating state of microwave power transistors, which can fully characterize their performance under specific operating conditions. By refining the output target into four key parameters, the model can focus on learning the nonlinear mapping relationship from S-parameters to the core operating point, avoiding redundant information interference, thereby achieving higher accuracy and more robust model training and generalization capabilities under limited data conditions.

[0009] Furthermore, in step two, the BiLSTM-LSTM structure establishes a fully connected layer between the BiLSTM layer and the LSTM layer. Its function is to receive the complete sequence output by the BiLSTM layer, refine and enhance the sequence through deep nonlinear transformation, and then pass it to the next LSTM layer with the same length. This allows the LSTM layer to focus on learning the long-term dependencies between these high-order features, which helps the model capture the nonlinear features in the data.

[0010] The beneficial effects of the above-mentioned further scheme are as follows: This invention achieves multi-level refinement and enhancement of the original temporal features by performing deep nonlinear transformation on the serialized features output by BiLSTM in the fully connected layer, which significantly improves the discriminativeness and representational ability of the feature space and provides more informative input for subsequent temporal modeling; secondly, this design makes the responsibilities of each layer of the model clear and the division of labor collaborative, effectively solving the optimization conflict that may exist in the model when local feature interaction and global temporal modeling, and improving the training stability and convergence efficiency of the overall model.

[0011] Furthermore, the MSE in step five is the average of the squared errors between the model's predicted values ​​and the actual values, which can keenly reflect abnormal errors in the prediction; R 2 It evaluates the model's ability to explain data variability and describes the model's goodness of fit; MAE is the average absolute error between predicted and true values, used to evaluate the model's stability. The formulas for calculating these three are as follows: , , , in, It is the true value of the i-th sample. It is the predicted value of the i-th sample. is the mean of the true values, and n is the total number of samples.

[0012] The beneficial effects of the above-mentioned further solutions are: the present invention comprehensively adopts a multi-dimensional evaluation system constructed by MSE, R² and MAE, and conducts rigorous quantitative evaluation of model performance from three complementary perspectives, effectively avoiding the limitations of a single index, realizing closed-loop verification of model performance "convergence accuracy - goodness of fit - practical error", and making the model evaluation results comprehensive and reliable. Attached Figure Description

[0013] Figure 1 This invention provides a reverse modeling process for GaAs pHEMT S-parameters based on BiLSTM-LSTM. Figure 2 This is the BiLSTM-LSTM model structure of the present invention; Figure 3 The above are the GaAs pHEMT f-fitting results based on the BiLSTM-LSTM model of this invention; Figure 4 The present invention relates to GaAs pHEMT I based on the BiLSTM-LSTM model. gs Fitting results; Figure 5 The present invention relates to GaAs pHEMT V based on the BiLSTM-LSTM model. ds Fitting results; Figure 6 The present invention relates to GaAs pHEMT V based on the BiLSTM-LSTM model. gs Fitting results; Detailed Implementation

[0014] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0015] This invention provides an S-parameter inverse modeling method for gallium arsenide pseudomorphic high electron mobility transistors (GaAspHEMTs) based on BiLSTM-LSTM technology, such as... Figure 1 As shown, the reverse modeling process of GaAs pHEMT S-parameters based on BiLSTM-LSTM includes five steps: data preparation, model design, hyperparameter setting, model training and optimization, and model evaluation.

[0016] Step one, data preparation, involves selecting GaAs pHEMTs within a frequency (f) range of 1MHz–30MHz, and experimentally collecting data at different gate currents (Ig). gs ), drain voltage (Vds ) and gate voltage (V gs The real and imaginary parts of the S-parameters under the corresponding conditions, i.e., [ReS 11 ImS 11 ReS 21 ImS 21 ReS 12 ImS 12 ReS 22 ImS 22 ] as input variables, with the corresponding frequencies and bias conditions [f, I gs V ds V gs As the output variable, a total of 2700 sets of data were used. The dataset was then divided into training and test sets in a 4:1 ratio. To improve the robustness of the model, Gaussian white noise with a standard deviation of 0.01 was added to the training set to double its sample size. Finally, the data was normalized to the [0, 1] interval and converted into the cell array required by BiLSTM.

[0017] Step two involves designing the model by, after determining the input and output data, using methods such as... Figure 2 The BiLSTM-LSTM model structure shown has an input layer consisting of 8 neurons corresponding to 8 input vectors [ReS]. 11 ImS 11 ReS 21 ImS 21 ReS 12 ImS 12 ReS 22 ImS 22 The BiLSTM layer consists of 20 units and outputs a complete sequence. The fully connected layer consists of 32 neurons and performs feature transformation on the data output by the BiLSTM layer. The LSTM layer consists of 10 units and is used to capture the final state. Finally, the output of the four neurons in the output layer is mapped to four output target values ​​[f, I]. gs V ds V gs ].

[0018] Step 3 sets the hyperparameters to use the Adam optimizer, which has the ability to adaptively adjust the learning rate, to help alleviate the gradient vanishing / exploding problem during training. The number of iterations is set to 30, the batch size to 64, and the initial learning rate and learning rate descent factor are set to 0.01 and 0.05, respectively.

[0019] Step four, model training and optimization, involves continuously monitoring the mean squared error (MSE) of the model during training. If the MSE value drops below 0.01, the model evaluation phase begins. The trained model is then used to predict the test set, and the normalized prediction results are denormalized to restore them to their original physical dimensions for quantitative performance analysis. Otherwise, the process returns to step two to optimize and adjust the BiLSTM-LSTM structure and retrain.

[0020] The fitting results of GaAs pHEMT f based on the BiLSTM-LSTM model are as follows: Figure 3 As shown, its R 2 The value is 0.9972; GaAs pHEMT I based on the BiLSTM-LSTM model gs The fitting results are as follows Figure 4 As shown, its R 2 The value is 0.9789; GaAs pHEMT V based on the BiLSTM-LSTM model ds The fitting results are as follows Figure 5 As shown, its R 2 The value is 0.9908; GaAspHEMT V based on the BiLSTM-LSTM model gs The fitting results are as follows Figure 6 As shown, its R 2 The value was 0.9933, and experiments showed that the predicted curve closely matched the true value. This was achieved through the use of MSE, MAE, and R... 2 The model performance was evaluated using the regression index, and the performance evaluation results for the modeling data are shown in Table 1. The best MSE is I. gs and V ds All are 0.0001, best R 2 The optimal MAE is for f = 0.9972 and for V. gs With a value of 0.0049, the model achieved excellent performance on the test set after comprehensive evaluation. The above results fully verify that the invention can accurately and reliably complete the reverse modeling of GaAs pHEMT transistors from S-parameters to operating point characteristic parameters.

[0021] Table 1 Performance evaluation results of the modeling data

[0022] In summary, this inverse modeling method for S-parameters of microwave power transistors based on BiLSTM-LSTM technology provides an important reference for small-signal modeling research and circuit design of microwave power transistors.

[0023] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for inverse modeling of S-parameters of microwave power transistors based on BiLSTM-LSTM technology, characterized in that, It includes five steps: data preparation, model design, hyperparameter setting, model training and optimization, and model evaluation. The data preparation in step one involves obtaining the dataset through experimental measurements, where the input variables are the S-parameters of GaAs pHEMT. ReS 11 , ImS 11 , ReS 21 , ImS 21 , ReS 12 , ImS 12 , ReS 22 , ImS 22 The output variables are composed of the corresponding microwave power transistor performance characteristics parameters. X = [ X 1, X 2, X 3, X The dataset is then randomly divided into training and testing sets in a 4:1 ratio, and the data is normalized and preprocessed. Finally, it is transformed into a cell array suitable for training temporal networks. The model design in step two involves creating a BiLSTM-LSTM structure using the layerGraph function in the Matlab deep learning framework after determining the input and output data, and then setting the number of neurons and initial weights in the neural network input layer, hidden layer, and output layer in sequence. The hyperparameters in step three are set as follows: the maximum number of iterations is set to 30, the initial learning rate and the learning rate decrease factor are 0.01 and 0.05 respectively, and then the Adam optimizer is used to update the weights. The fourth step, model training and optimization, is as follows: when the training process satisfies the condition that the loss function value no longer decreases significantly or the number of iterations reaches the maximum value of 30, the training ends. At the same time, the mean square error (MSE) value is continuously monitored during model training. If the MSE value drops below 0.01, the process proceeds to the fifth step, model evaluation. Otherwise, the process returns to the second step, model design, adjusts the number of hidden layer neurons, and retrains. The model evaluation in step five uses MSE, Mean Absolute Error (MAE), and Coefficient of Determination (R-Square, R0). 2 () is used as a regression index to evaluate the model's performance.

2. The method for inverse modeling of microwave power transistor S-parameters based on BiLSTM-LSTM technology according to claim 1, characterized in that, The characteristic parameters of step one X = [ X 1, X 2, X 3, X 4], of which X 1 represents frequency (f). X 2 is the gate current (I) gs ), X 3 represents the drain voltage (V) ds ), X 4 represents the gate voltage (V) gs ).

3. The method for inverse modeling of microwave power transistor S-parameters based on BiLSTM-LSTM technology according to claim 1, characterized in that, The BiLSTM-LSTM structure in step two establishes a fully connected layer between the BiLSTM layer and the LSTM layer. Its function is to receive the complete sequence output by the BiLSTM layer, refine and enhance the sequence through deep nonlinear transformation, and then pass it to the next LSTM layer with the same length. This allows the LSTM layer to focus on learning the long-term dependencies of high-order features, which helps the model capture nonlinear features in the data.

4. The method for inverse modeling of microwave power transistor S-parameters based on BiLSTM-LSTM technology according to claim 1, characterized in that, The MSE in step five is the average of the squared errors between the model's predicted values ​​and the actual values, which can sensitively reflect abnormal errors in the prediction; R 2 It evaluates the model's ability to explain data variability and describes the model's goodness of fit; MAE is the average absolute error between predicted and true values, used to evaluate the model's stability. The formulas for calculating these three are as follows: ， ， ， in, It is the true value of the i-th sample. It is the predicted value of the i-th sample. is the mean of the true values, and n is the total number of samples.

Images