DNN-based large signal model modeling method, system and device and storage medium

Abstract

The invention relates to a DNN-based large signal model modeling method, system and device and a storage medium, and relates to the technical field of integrated circuits, the method is based on a GaN microwave device comprising a GaN transistor, and the method comprises the following steps: carrying out a large signal test on the GaN transistor to obtain traveling wave characteristics; processing input layer characteristics and output layer characteristics in the traveling wave characteristics to obtain input layer data and output layer data; extracting real part and imaginary part data, and performing DNN training to obtain a high-power region deep neural network model; amplitude and phase data are extracted, and a low-power region deep neural network model is obtained through DNN training; training traveling wave characteristics according to the high-power area deep neural network model and the low-power area deep neural network model to obtain training traveling wave characteristics, and converting the training traveling wave characteristics into voltage and current characteristics; and a large signal behavior model is constructed according to the voltage and current characteristics. The method has the technical effect that the accuracy of the large-signal modeling model of the GaN microwave device is improved.

Description

Technical Field

[0001] This application relates to the field of integrated circuit technology, and particularly to a method, system, device, and storage medium for modeling a large-signal model based on DNN. Background Art

[0002] DNN is short for Deep Neural Network, which is a multi-layer neural network structure designed to enhance the model's expressive power and ability to learn complex features by stacking multiple hidden layers.

[0003] With the development of today's infrastructure, communication industry, and aerospace technology, higher requirements are put forward for the microwave systems of devices: miniaturization, high temperature resistance, radiation resistance, high power, ultra-high frequency, suitable for working in harsh environments, etc. Wide-bandgap semiconductor materials and devices represented by GaN and SiC have become research hotspots. It is of great significance to research and develop high-performance semiconductor materials and devices that can work at higher frequencies and have greater power.

[0004] With the continuous improvement of the quality of epitaxial materials and the continuous improvement of device processes, the development of AlGaN / GaN HEMT devices has been very rapid. Due to its spontaneous polarization and piezoelectric polarization effects, the AlGaN / GaN heterojunction can generate a two-dimensional electron gas with very high mobility, which can be controlled by the gate voltage. In recent years, the characteristic indicators of the devices have developed rapidly, especially the microwave power characteristics of AlGaN / GaN HEMT devices. The output power and power density of the devices have been greatly improved. In addition to manufacturing process technology and device characteristics, the modeling of AlGaN / GaN HEMT has also been a key research point. Due to its outstanding applications in the radio frequency and microwave fields, the research on AlGaN / GaN HEMT device models has always been an important part of the device research field.

[0005] Device models include small-signal models and large-signal models. For small-signal models, certain progress has been made in small-signal modeling both at home and abroad, and the S-parameter method is the most widely used.

[0006] However, large-signal modeling has always been a difficult point in the analysis of microwave power devices. Commonly used GaN HEMT device models can generally be divided into: empirical analytical models, physical models, table-based models, and behavioral models. Due to reasons such as the mutability of the performance curves of existing model transistors in the low-power and high-power regions, it seriously affects the accuracy of the large-signal modeling model, resulting in relatively low model accuracy. Summary of the Invention

[0007] To improve the accuracy of the large-signal modeling of GaN microwave devices, this application provides a large-signal model modeling method, system, device, and storage medium based on DNN.

[0008] In the first aspect, this application provides a large-signal model modeling method based on DNN, adopting the following technical solutions:

[0009] Perform large-signal testing on the GaN transistor to obtain traveling-wave characteristics, which include input-layer characteristics and output-layer characteristics;

[0010] Process the input-layer characteristics to obtain input-layer data, and process the output-layer characteristics to obtain output-layer data;

[0011] Extract the real and imaginary parts of the input-layer data and the output-layer data, and use DNN to train based on the real and imaginary parts to obtain a high-power region deep neural network model;

[0012] Extract the amplitude and phase data of the input-layer data and the output-layer data, and use DNN to train based on the amplitude and phase data to obtain a low-power region deep neural network model;

[0013] Train the traveling-wave characteristics based on the high-power region deep neural network model and the low-power region deep neural network model to obtain trained traveling-wave characteristics, and convert the trained traveling-wave characteristics into voltage-current characteristics;

[0014] Construct a large-signal behavior model based on the voltage-current characteristics.

[0015] Through the above technical solutions, using piecewise modeling can simultaneously fit the large-signal characteristics in the low-power and high-power ranges. Using a deep neural network to construct a behavior model highlights the weights of the traveling-wave parameters in the high-power region and the low-power region, enhances the learning sensitivity of the neural network, and improves the accuracy of the large-signal modeling model.

[0016] In a specific feasible implementation, the traveling-wave characteristics obtained by performing large-signal testing on the GaN transistor include:

[0017] Obtain the fundamental wave conditions and harmonic conditions for testing;

[0018] Perform input power scanning on the GaN transistor through a vector network analyzer under the fundamental wave conditions and the harmonic conditions to obtain a scanning result;

[0019] Read the scanning result through a coupler to obtain traveling-wave characteristics.

[0020] Through the above technical solution, large-signal testing is carried out by using vector load-pull testing under the conditions of a fixed fundamental wave and harmonic scanning range, and more accurate traveling-wave testing of transistors can be obtained.

[0021] In a specific feasible implementation, the processing of the input layer characteristics to obtain input layer data and the processing of the output layer characteristics to obtain output layer data include:

[0022] Calculating the target phase of the fundamental wave incident wave at the input port;

[0023] Normalizing the remaining incident waves according to the target phase to obtain input layer data;

[0024] Normalizing the reflected wave according to the target phase to obtain output layer data.

[0025] Through the above technical solution, the dynamically changing input layer characteristics and output layer characteristics are converted into time-invariant data, eliminating the periodic changes in the phase data, making it easier to compare and analyze.

[0026] In a specific feasible implementation, the extraction of the real part data and the imaginary part data of the input layer data and the output layer data includes:

[0027] Extracting the target amplitude of the fundamental wave incident wave at the input port, extracting the real part data and the imaginary part data of the input layer data, and extracting the real part data and the imaginary part data of the output layer data;

[0028] The use of the DNN to train the high-power region deep neural network model according to the real part data and the imaginary part data includes:

[0029] Normalizing the target amplitude, the real part data and the imaginary part data of the input layer data, and the real part data and the imaginary part data of the output layer data to obtain high-power region training data;

[0030] Importing the high-power region training data into a preset first neural network for training and calculating the first loss value;

[0031] Judging whether the first loss value is less than a preset value;

[0032] If the first loss value is less than the preset value, stop training and obtain the high-power region deep neural network model according to the trained first neural network.

[0033] Through the above technical solution, a high-power region deep neural network model is trained by extracting real part data and imaginary part data for partition processing, improving the accuracy of the large-signal modeling model.

[0034] In a specific feasible implementation, the extraction of the amplitude data and phase data of the input layer data and the output layer data includes:

[0035] Extract the dB amplitude of the fundamental incident wave at the input port, where the unit of the dB amplitude is dB, extract the amplitude data and phase data of the input layer data, and extract the amplitude data and phase data of the output layer data;

[0036] The training of the low-power region deep neural network model using DNN based on the amplitude data and the phase data includes:

[0037] Normalize the dB amplitude, the amplitude data and phase data of the input layer data, and the amplitude data and phase data of the output layer data to obtain low-power region training data;

[0038] Import the low-power region training data into a preset second neural network for training and calculate the second loss value;

[0039] Determine whether the second loss value is less than a second preset value;

[0040] If the second loss value is less than the second preset value, stop training and obtain the low-power region deep neural network model according to the trained second neural network.

[0041] Through the above technical solution, by extracting the amplitude data and phase data, a low-power region deep neural network model is trained in a partitioned manner, and targeted processing is performed in the partition to improve the accuracy of the large-signal modeling model.

[0042] In a specific feasible implementation, the training of the traveling wave characteristics according to the high-power region deep neural network model and the low-power region deep neural network model to obtain the trained traveling wave characteristics includes:

[0043] Set the determination point for the high and low power regions;

[0044] Construct a partitioned deep neural network model according to the determination point, the high-power region deep neural network model, and the low-power region deep neural network model;

[0045] Train the traveling wave characteristics according to the partitioned deep neural network model and perform anti-normalization processing to obtain the trained traveling wave characteristics.

[0046] Through the above technical solution, the trained high-power region deep neural network model and low-power region deep neural network model are used to generate an overall partitioned deep neural network model by setting the high and low power region determination points and a single-pole double-throw switch, and the traveling wave characteristics of both the high-power and low-power regions can be trained with a large order of magnitude simultaneously.

[0047] In a specific feasible implementation, after constructing the large-signal behavior model according to the voltage-current characteristics, the following steps are further included:

[0048] Receiving and verifying the traveling-wave characteristics, where the verified traveling-wave characteristics include the actual data of the traveling-wave characteristics;

[0049] Simulating the verified traveling-wave characteristics using the large-signal behavior model to obtain simulation data:

[0050] Judging whether the error between the actual data and the simulation data is within the allowable range;

[0051] If the error is within the allowable range, it is determined that the large-signal behavior model can be put into use;

[0052] If the error is not within the allowable range, it is determined that the large-signal behavior model cannot be put into use.

[0053] Through the above technical solution, the obtained large-signal behavior model is simulated and compared to verify the function of the large-signal behavior model, ensuring the accuracy of the large-signal behavior model.

[0054] In a second aspect, the present application provides a large-signal model modeling system based on DNN, adopting the following technical solution: The system is based on a GaN microwave device including a GaN transistor, and the system includes:

[0055] A traveling-wave characteristic acquisition module, configured to perform large-signal tests on the GaN transistor to obtain traveling-wave characteristics, where the traveling-wave characteristics include input layer characteristics and output layer characteristics;

[0056] A traveling-wave characteristic processing module, configured to process the input layer characteristics to obtain input layer data, and process the output layer characteristics to obtain output layer data;

[0057] A high-power region model training module, configured to extract the real part data and imaginary part data of the input layer data and the output layer data, and use DNN to train a high-power region deep neural network model according to the real part data and the imaginary part data;

[0058] A low-power region model training module, configured to extract the amplitude data and phase data of the input layer data and the output layer data, and use DNN to train a low-power region deep neural network model according to the amplitude data and the phase data;

[0059] A voltage-current characteristic conversion module, configured to train the traveling-wave characteristics according to the high-power region deep neural network model and the low-power region deep neural network model to obtain trained traveling-wave characteristics, and convert the trained traveling-wave characteristics into voltage-current characteristics;

[0060] A behavior model construction module, configured to construct a large-signal behavior model according to the voltage-current characteristics.

[0061] In a third aspect, the present application provides a computer device, adopting the following technical solution: including a memory and a processor, and a computer program capable of being loaded and executed by the processor, such as the above-mentioned large-signal model modeling method based on DNN, is stored on the memory.

[0062] In a fourth aspect, the present application provides a computer-readable storage medium, adopting the following technical solution: storing a computer program capable of being loaded and executed by the processor for the above-mentioned large-signal model modeling method based on DNN.

[0063] In summary, the present application has the following beneficial technical effects:

[0064] (1) By using piecewise modeling, the large-signal characteristics in both low-power and high-power ranges can be fitted simultaneously. Constructing a behavior model using a deep neural network highlights the weights of the traveling-wave parameters in the high-power region and the low-power region, enhances the learning sensitivity of the neural network, and improves the accuracy of the large-signal modeling model.

[0065] (2) Constructing a behavior model using a deep neural network processes the input layer and the output layer through complex real part / imaginary part and amplitude / phase algorithms, overcomes the fitting limitations of the neural network for data with mutations, and can train the traveling-wave characteristics in both high-power and low-power regions with a large order of magnitude, thereby expanding the dynamic range of the model.

[0066] (3) Adding frequency and voltage variables to the input layer of the neural network and combining the method of partition algorithm modeling enable the behavior model to have the ability to fit broadband and multi-bias conditions.

[0067] (4) The model constructed in the present application can not only be modeled at the transistor level, but also be accurately modeled at the integrated circuit module level to complete system design; at the same time, it can also be compatible with various integrated circuit design software, facilitating circuit and system engineers to call and simulate without additional hardware and software drivers. Description of the Drawings

[0068] Figure 1 is the flowchart in the embodiment of the present application

[0069] Figure 2 is the cross-sectional view of the GaN device under test.

[0070] Figure 3 is the example diagram of the fundamental wave and second harmonic scanning at the output port.

[0071] Figure 4 is the input power dynamic scanning range diagram.

[0072] Figure 5 It is the traveling wave characteristic of the real / imaginary part of a complex number in the linear coordinate.

[0073] Figure 6 It is the traveling wave characteristic of the real / imaginary part of a complex number in the logarithmic coordinate.

[0074] Figure 7 It is the traveling wave characteristic of the amplitude / phase mode in the linear coordinate.

[0075] Figure 8 It is the traveling wave characteristic of the amplitude / phase mode in the logarithmic coordinate.

[0076] Figure 9 It is the simulation curve of the model in the real / imaginary part format of a complex number.

[0077] Figure 10 It is the simulation curve of the model in the amplitude / phase format.

[0078] Figure 11 It is the simulation result of the partitioned deep neural network behavior model.

[0079] Figure 12 It is the framework of the deep neural network model.

[0080] Figure 13 The simulation curve for verifying the traveling wave characteristic by the large-signal behavior model simulation.

[0081] Figure 14 It is the simulation of the GaN power amplifier module.

[0082] Figure 15 It is the structural block diagram in the embodiment of the present application.

[0083] Reference signs: 1501, traveling wave characteristic acquisition module; 1502, traveling wave characteristic processing module; 1503, high-power region model training module; 1504, low-power region model training module; 1505, voltage-current characteristic conversion module; 1506, behavior model construction module. Detailed implementation manners

[0084] The following is a further detailed description of the present application with reference to the Figures 1 - 15 accompanying drawings.

[0085] The embodiment of the present application discloses a large-signal model modeling method based on DNN, which is used to improve the accuracy of the large-signal modeling model of GaN microwave devices.

[0086] DNN is the abbreviation of Deep Neural Network. It is a multi-layer neural network structure designed to enhance the model's expressive ability and the ability to learn complex features by stacking multiple hidden layers.

[0087] With the development of today's infrastructure, communication industry, and aerospace technology, higher and higher requirements are put forward for the microwave systems of equipment: miniaturization, high temperature resistance, radiation resistance, high power, ultra-high frequency, suitable for working in harsh environments, etc. Wide bandgap semiconductor materials and devices represented by GaN and SiC have become the focus of research. It is of great significance to research and develop high-performance semiconductor materials and devices that can work at higher frequencies and have greater power.

[0088] With the continuous improvement of the quality of epitaxial materials and the continuous improvement of device processes, the development of AlGaN / GaN HEMT devices has been very rapid. Due to its spontaneous polarization and piezoelectric polarization effects, the AlGaN / GaN heterojunction can generate a two-dimensional electron gas with a very high mobility, which can be controlled by the gate voltage. In recent years, the characteristic indexes of the devices have developed rapidly, especially the microwave power characteristics of AlGaN / GaN HEMT devices. The output power and power density of the devices have been greatly improved. In addition to manufacturing process technology and device characteristics, the modeling of AlGaN / GaN HEMT has also been the focus of research. Due to its outstanding applications in the radio frequency and microwave fields, the research on AlGaN / GaN HEMT device models has always been an important part of the device research field.

[0089] Device models include small-signal models and large-signal models. For small-signal models, whether at home or abroad, certain progress has been made in small-signal modeling, and the S-parameter method is the most widely used.

[0090] Large-signal modeling has always been the difficulty in the analysis of microwave power devices. Commonly used GaN HEMT device models can generally be divided into the following types:

[0091] I. Empirical analytical models, also called equivalent circuit models. That is, for different devices fabricated by actual processes, corresponding testing and parameter extraction techniques are adopted, and various device models required for fabricating monolithic circuits are obtained by parameter fitting and optimization methods, such as the EEHEMT model and the Angelov model. The model is divided into two parts: parasitic parameters and intrinsic parameters. The intrinsic capacitance and current models fit the test data through analytical formulas. Its advantages are that the functional relationship and algorithm are simple, and it is very suitable for large-signal radio frequency amplifiers. However, these models all require the extraction of a certain amount of fitting parameters, and the higher the model accuracy, the more fitting parameters are required, thus increasing the complexity of the expression. Often, a trade-off needs to be made between the complexity and accuracy of the model. Since the formula lacks physical meaning, it is impossible to accurately fit the non-ideal effects found in the characterization process.

[0092] II. Physical Models. Commonly used physical models include the ASM model based on surface potential and the MVSG model based on charge. Compared with the equivalent circuit model, the current and charge expressions of the physical model are modeled based on the device's structural parameters and semiconductor physics formulas, so it can accurately fit the DC and AC characteristics of the device. However, for integrated circuit design companies without a processing line, it is difficult to obtain the physical parameters of the device structure, and the material parameters of the epitaxial layer are even unknown. Therefore, the use of physical models is greatly restricted. In addition, due to the coupling of its DC parameters and AC parameters, it brings challenges to parameter extraction and optimization in modeling.

[0093] III. Table-based Model, also known as the data-based model, is a model established on the basis of a large number of tests. Because it is a true reflection of the device's electrical parameters, it is more closely combined with the processing line and closer to the actual device characteristics. However, the flexibility and practicality of the table-based model are worse than those of the empirical analytical model. Accurate modeling requires measuring more data, which reduces its scalability and increases the difficulty of the statistical average model.

[0094] IV. Behavioral Model. By black-boxing the transistor and measuring the large-signal characteristics in the frequency domain or time domain to obtain the characteristics of the device ports, a behavioral model is established, which has the highest accuracy like the table-based model.

[0095] Commonly used behavioral models include the X-parameter model and the Cardiff model. Among them, the research and application of the Cardiff model are more sufficient. The basic principle of the Cardiff model is the harmonic superposition principle and the harmonic mixing principle. Model coefficients are extracted by setting a large-signal operating point, and then the performance of the model at this large-signal operating point is calculated by the look-up table method. However, this method is difficult to achieve wide bandwidth and multi-bias conditions, and there is no accurate formula guidance for interpolation between test points in the look-up table method. Therefore, the accuracy cannot be guaranteed. The other is to establish a behavioral model based on artificial neural networks in the frequency-domain traveling-wave domain. However, due to the mutability of the performance curves of transistors in the low-power and high-power regions, artificial neural networks cannot fit this operating mode. Therefore, it reduces the dynamic range of its input power and limits the application of its multi-bias conditions.

[0096] By black-boxing the transistor through establishing a behavioral model, problems such as the low accuracy of empirical models and physical models, unknown physical parameters, and difficulty in parameter extraction can be solved. However, there are still some problems. For example, due to reasons such as the mutability of the performance curves of transistors in the low-power and high-power regions, it will seriously affect the accuracy of the large-signal modeling model, resulting in a relatively low model accuracy.

[0097] Therefore, this application proposes a large-signal model modeling method based on DNN, and uses this method to improve the accuracy of the large-signal modeling model of GaN microwave devices.

[0098] As shown in Figure 1 the method includes:

[0099] S10. Conduct large-signal testing on the GaN transistor to obtain traveling-wave characteristics, which include input-layer characteristics and output-layer characteristics.

[0100] Specifically, for the traveling-wave domain in the frequency domain, conduct large-signal testing on the GaN transistor to obtain traveling-wave characteristics, frequency, and voltage. The traveling-wave characteristics include input-layer characteristics and output-layer characteristics. The input-layer characteristics include the incident wave, and the output-layer characteristics include the reflected wave.

[0101] S20. Process the input-layer characteristics to obtain input-layer data, and process the output-layer characteristics to obtain output-layer data.

[0102] Specifically, normalize the input-layer characteristics and output-layer characteristics, and perform phase normalization processing on the traveling-wave characteristics through the incident wave of the fundamental wave at the input port to convert them into time-invariant data, obtaining input-layer data and output-layer data.

[0103] S30. Extract the real-part data and imaginary-part data of the input-layer data and output-layer data, and use the DNN to train based on the real-part data and imaginary-part data to obtain a high-power region deep neural network model.

[0104] Specifically, the performance curves of the transistor in the low-power and high-power regions have mutations, so partitioned training is carried out. In the high-power region of the traveling-wave characteristics, complex real part / imaginary part is used for partitioned deep neural network training, that is, extract the real-part data and imaginary-part data of the input-layer data and output-layer data in the high-power region, and use the DNN to train based on the real-part data and imaginary-part data to obtain a high-power region deep neural network model.

[0105] S40. Extract the amplitude data and phase data of the input-layer data and output-layer data, and use the DNN to train based on the amplitude data and phase data to obtain a low-power region deep neural network model.

[0106] Specifically, the performance curves of the transistor in the low-power and high-power regions have mutations, so partitioned training is carried out. In the low-power region of the traveling-wave characteristics, amplitude / phase is used for partitioned deep neural network training, that is, extract the amplitude data and phase data of the input-layer data and output-layer data in the low-power region, and use the DNN to train based on the amplitude data and phase data to obtain a low-power region deep neural network model.

[0107] S50. Train the traveling-wave characteristics based on the high-power region deep neural network model and the low-power region deep neural network model to obtain trained traveling-wave characteristics, and convert the trained traveling-wave characteristics into voltage-current characteristics.

[0108] Specifically, the trained traveling wave characteristics of the partitioned deep neural network model are converted into voltage and current characteristics to facilitate the establishment of the behavioral model.

[0109] S60. A large-signal behavioral model is constructed based on the voltage and current characteristics.

[0110] Specifically, in the ADS circuit design software, the FDD control is used to define the port voltage and current, and the behavioral model is constructed based on the voltage and current.

[0111] By using piecewise modeling, the large-signal characteristics in the low-power and high-power ranges can be fitted simultaneously. Constructing the behavioral model using a deep neural network highlights the weights of the traveling wave parameters in the high-power region and the low-power region, enhances the learning sensitivity of the neural network, and improves the accuracy of the large-signal modeling model.

[0112] In one embodiment, to improve the accuracy of the large-signal modeling model of the GaN microwave device, the step of obtaining the traveling wave characteristics by performing large-signal testing on the GaN transistor can be specifically implemented as follows:

[0113] First, the fundamental wave conditions and harmonic conditions for testing are obtained. Specifically, as shown in the GaN device Figure 2 which includes a fixed frequency, voltage, and input impedance. To obtain sample data for training, by changing the output impedance tuner, a scan is performed in the range near the optimal fundamental wave impedance point of power and efficiency, and the fundamental wave impedance point and the second harmonic impedance point are found to obtain the fundamental wave conditions and harmonic conditions for testing.

[0114] Next, an input power scan of the GaN transistor is performed under the fundamental wave conditions and harmonic conditions using a vector network analyzer to obtain a scan result. Specifically, as shown in Figure 3 , for each scanned fundamental wave impedance point, the second harmonic impedance point is scanned from 0° to 360° with a fixed phase step in the range where the magnitude of the reflection coefficient is greater than 0.8 to obtain the scan result.

[0115] Then, the traveling wave characteristics are obtained by reading the scan result through a coupler. Specifically, under the fixed fundamental wave and harmonic conditions, an input power scan is performed from the linear region to the non-linear region of gain compression. The large-signal characteristics of the transistor have obvious mutations in the low-power region and the high-power region. As shown in Figure 4 , the abscissa represents the power magnitude, and the ordinate represents the traveling wave characteristics. It can be seen that there are obvious mutations in the traveling wave characteristics in the low-power region and the high-power region shown in the figure. The traveling wave characteristics are obtained by reading the vector network analyzer through a coupler. The traveling wave characteristics include the input layer characteristics and the output layer characteristics. The input layer characteristics include the incident wave A ph , frequency, and voltage, and the output layer characteristics include the reflected wave B ph, where p represents the port index, p = 1 is the input port, p = 2 is the output port, h represents the harmonic index. When h = 0, it is the DC component; when h = 1, it is the fundamental wave f0 component; when h = 2, it is the second harmonic 2f0 component.

[0116] The large-signal test is carried out by using vector load-pull test under the conditions of fixed fundamental wave and harmonic scanning ranges, which can obtain more accurate traveling-wave characteristics of the transistor. At the same time, frequency and voltage variables are added to the input layer, and combined with the method of zoning algorithm modeling, the behavioral model has the ability to fit wide frequency and multi-bias conditions.

[0117] In one embodiment, in order to improve the accuracy of the large-signal modeling model of GaN microwave devices, the step of processing the input layer characteristics to obtain the input layer data and processing the output layer characteristics to obtain the output layer data can be specifically executed as follows:

[0118] First, calculate the target phase of the fundamental wave incident wave at the input port. Specifically, calculate the phase Angle(A 11 ) of the fundamental wave incident wave A at the input port. 11 )

[0119] Then, normalize the remaining incident waves according to the target phase to obtain the input layer data, and normalize the reflected waves according to the target phase to obtain the output layer data. Specifically, normalize the input layer data except A 11 and the B 11 in the output layer data according to Angle(A ph ). The formula is as follows: ph

[0120] A ph_norm = A ph / (cos(h * Angle(A 11 )) + i * sin(h * Angle(A 11 )))

[0121] B ph_m = B ph / (cos(h * Angle(A 11 )) + i * sin(h * Angle(A 11 )))

[0122] Where A ph_norm and B ph_m are the normalized input layer data (incident wave component) and output layer data (reflected wave component), and h > 0.

[0123] It should be noted that the normalization process is to convert the traveling-wave characteristics into time-invariant data. Therefore, this step is for non-DC components, and the additional limiting condition is h > 0. When h = 0, for example, B 20The 2-port current values obtained from the test, with the unit of A, B 20 are fixed values and do not require normalization processing.

[0124] Convert the characteristics of the dynamic input layer and output layer into time-invariant data, eliminate the periodic variation of the phase data, and make it easier to compare and analyze.

[0125] In one embodiment, in order to improve the accuracy of the large-signal modeling model of GaN microwave devices, the step of extracting the real and imaginary parts of the input layer data and output layer data can be specifically executed as follows:

[0126] Extract the target amplitude of the fundamental incident wave at the input port, extract the real and imaginary parts of the input layer data, and extract the real and imaginary parts of the output layer data. Specifically, A 11 's phase is used for time-invariant normalization. Therefore, A 11 has only one variable. In the first step, extract the amplitude of A 11 , and the formula is as follows:

[0127] A 11_ag = abs(A 11 )

[0128] In the second step, extract the real and imaginary parts of A ph_m and B ph_rm , and the formula is as follows:

[0129] A ph_ = real(A ph_orm )

[0130] B ph_ = real(B ph_m )

[0131] A ph_ag = imag(A ph_orm )

[0132] B ph_ = imag(B ph_m )

[0133] where A ph_ is the real part component of the incident wave, B ph_ is the real part component of the reflected wave, A ph_ is the imaginary part component of the incident wave, and B ph_ is the imaginary part component of the reflected wave. As Figure 5 and Figure 6Shows the real and imaginary parts of the components of the traveling wave, where abs(Re(A21)) refers to the absolute value of the real part data of A21, abs(Im(A21)) is the absolute value of the imaginary part data of A21, abs(Re(B21)) refers to the absolute value of the real part data of B21, abs(Im(B21)) is the absolute value of the imaginary part data of B21, and the abscissa represents the power magnitude.

[0134] Figure 5 The ordinate of is the traveling wave characteristic under linear coordinates, Figure 6 The ordinate of Figure 6 is the traveling wave characteristic under logarithmic coordinates. From Figure 5 it can be seen that there is no obvious difference between the low-power region and the high-power region under linear coordinates, but in Figure 6 it can be seen that the difference between the low-power and high-power regions under logarithmic coordinates exceeds four orders of magnitude and is obvious in the high-power region. Therefore, the complex real part / imaginary part is used for the partition depth neural network training in the high-power region of the traveling wave characteristic.

[0135] The step of using the real part data and the imaginary part data to train the depth neural network model in the high-power region by DNN can be specifically executed as follows:

[0136] First, normalize the target amplitude, the real part data and the imaginary part data of the input layer data, and the real part data and the imaginary part data of the output layer data to obtain the high-power region training data. Specifically, normalize the extracted data of the input layer and the extracted data of the output layer to obtain the high-power region training data. The normalization formula is as follows:

[0137]

[0138] where, X i is the i-th value input with the change of the input power, X min is the minimum value of this parameter, X max is the maximum value of this parameter, and X i_ is the normalized value.

[0139] Then, import the high-power region training data into a preset first neural network for training and calculate the first loss value. Specifically, the first neural network is a depth neural network with three hidden layers, and each hidden layer has 16 neurons. Import the high-power region training data into the preset first neural network for training to obtain the simulation value. The actual value of the sample is the test value. Calculate the loss value according to the simulation value and the test value. The formula for calculating the loss value is as follows:

[0140]

[0141] where, y i is the simulation value, yi_ Is a test value.

[0142] Next, determine whether the first loss value is less than the preset value. If the first loss value is less than the preset value, stop training and obtain the high-power region deep neural network model according to the trained first neural network. Specifically, during the training process, calculate the value of the loss function. If the value of the loss function is less than the preset value, we stop training and consider that this first neural network can be used to train the traveling wave characteristics of the high-power region, and obtain a high-power region deep neural network model that can be put into use according to the trained first neural network; if it is greater than the preset value, use the stochastic gradient descent optimization method to optimize the first neural network and then continue training until the training cycle reaches a certain number of times.

[0143] Train the high-power region deep neural network model by extracting the real part data and the imaginary part data for partition processing, which improves the accuracy of the large-signal modeling model.

[0144] In one embodiment, in order to improve the accuracy of the large-signal modeling model of the GaN microwave device, the step of extracting the amplitude data and phase data of the input layer data and the output layer data can be specifically implemented as follows:

[0145] Extract the dB amplitude of the fundamental incident wave at the input port. The unit of the dB amplitude is dB. Extract the amplitude data and phase data of the input layer data, and extract the amplitude data and phase data of the output layer data. Specifically, the 11 phase of A is used for time-invariant normalization. Therefore, A 11 has only one variable. In the first step, extract the 11 amplitude in dB. The formula is as follows:

[0146] A 11_mag_ = 20 log(abs(A 11 ))

[0147] where A 11_ag_ is the A in dB 11 amplitude magnitude.

[0148] In the second step, extract the amplitudes and phases of A ph_m and B ph_rm . The formula is as follows:

[0149]

[0150] A ph_phase_ad = atan2(A ph_ag , A ph_ ) B ph_ase_ad = atan2(B ph_ , B ph_ )

[0151] where A ph__ is the amplitude of the incident wave in dB, and B ph__ is the amplitude of the reflected wave in dB, and A ph__ is the phase of the incident wave in rad, and B ph__ is the phase of the reflected wave in rad. When h = 0, B 20_ = 20log(B 20 ), where B 20_ is the current component at port 2 in dB.

[0152] Figure 7 and Figure 8 show the amplitudes and phases of the components of the traveling wave, where abs(Mag_A21_dB) refers to the absolute value of the amplitude data of A21, abs(Phase_A21) is the absolute value of the phase data of A21, abs(Mag_B21_dB) refers to the absolute value of the amplitude data of B21, abs(Phase_B21) is the absolute value of the phase data of B21, and the abscissa represents the power level.

[0153] Figure 7 The ordinate of Figure 8 is the traveling wave characteristic in linear coordinates, Figure 7 and the ordinate of Figure 8 is the traveling wave characteristic in logarithmic coordinates. It can be seen from

[0154] that in linear coordinates, the low-power region and the high-power region do not differ significantly, but it can be seen from

[0155] that in logarithmic coordinates, the low-power region and the high-power region show a weight difference of about four orders of magnitude, and it is obvious in the low-power region. Therefore, amplitude / phase is used for partitioning depth neural network training in the low-power region of the traveling wave characteristic.

[0156]

[0157] where, X i is the i-th value input with the change of the input power, X min is the minimum value of this parameter, X max is the maximum value of this parameter, and X i_ is the normalized value.

[0158] Next, import the low-power region training data into a preset second neural network for training and calculate the second loss value. The second neural network is used to train the low-power region deep neural network model and is a deep neural network with three hidden layers, with 16 neurons in each hidden layer. Import the low-power region training data into the preset second neural network for training to obtain a simulation value, and the actual value of the sample is the test value. Calculate the loss value based on the simulation value and the test value. The formula for calculating the loss value is as follows:

[0159]

[0160] where y i is the simulation value, and y i_ is the test value.

[0161] Then, determine whether the second loss value is less than the second preset value. If the second loss value is less than the second preset value, stop training and obtain the low-power region deep neural network model according to the trained second neural network. Specifically, during the training process, calculate the value of the loss function. If the value of the loss function is less than the preset value, stop training and consider that this second neural network can be used to train the traveling wave characteristics of the low-power region. Obtain the low-power region deep neural network model that can be put into use according to the trained second neural network; if it is greater than the preset value, use the stochastic gradient descent optimization method to optimize the second neural network and then continue training until the training cycle reaches a certain number of times.

[0162] Train the low-power region deep neural network model by extracting amplitude data and phase data for partition training, and perform partition processing to improve the accuracy of the large-signal modeling model.

[0163] It should be noted that in the prior art, there is a situation where only based on real part data and imaginary part data, a deep neural network model is trained using DNN, and its simulation curve is as Figure 9 shown. The simulation curve has a good fitting effect only in the high-power region, and the simulation curve in the low-power region changes greatly and the simulation is inaccurate; if only based on amplitude data and phase data, a deep neural network model is trained using DNN, and its simulation curve is as Figure 10 shown. The simulation curve has a good fitting effect only in the low-power region, and the simulation curve in the high-power region changes greatly and the simulation is inaccurate. The technical solution of this application obtains a deep neural network model through partition training, and its simulation curve is as Figure 11 shown. The fitting effect of the simulation curve is not affected by the power level and has a good fitting effect.

[0164] In one embodiment, in order to improve the accuracy of the large-signal modeling model of GaN microwave devices, the step of obtaining the trained traveling-wave characteristics by training the traveling-wave characteristics according to the high-power region deep neural network model and the low-power region deep neural network model can be specifically performed as follows:

[0165] First, set the determination point for the high and low power regions. Specifically, set P in_ as the determination point for the high and low power regions.

[0166] Next, construct a partitioned deep neural network model according to the determination point, the high-power region deep neural network model, and the low-power region deep neural network model. Specifically, as mentioned above, the first neural network and the second neural network are deep neural networks with three hidden layers, and each hidden layer has 16 neurons. The specific structure of the constructed partitioned deep neural network model is as Figure 12 shown. The activation function of the partitioned deep neural network model is set as the Sigmoid function, and the formula is as follows:

[0167]

[0168] where z is the weighted function, and the formula is as follows:

[0169] z = ∑w i x i + b

[0170] where x u is the independent variable of the neural network, w u is the weight, and b is the bias.

[0171] The following formula shows the training variable settings for the input layer and the output layer:

[0172]

[0173] where, P in_ is the determination point for the high and low power regions. The single-pole double-throw switch function is realized through the if-else selection determination statement to achieve intelligent switching between high and low powers. DNN_L is the trained deep neural network behavior model for the low-power region, and DNN_H is the trained deep neural network behavior model for the high-power region. Freq and V gs are the frequency and the gate voltage respectively, where the dimension of the frequency is GHz and the dimension of the voltage is V.

[0174] Next, train the traveling-wave characteristics according to the partitioned deep neural network model and perform anti-normalization processing to obtain the trained traveling-wave characteristics. Specifically, after constructing the partitioned deep neural network model, perform anti-normalization processing to restore the traveling-wave characteristic format, and the trained traveling-wave characteristics of the trained partitioned neural network model can be obtained.

[0175] The trained deep neural network model for the high-power region and the deep neural network model for the low-power region are used to generate an overall partitioned deep neural network model by setting a high-low power region determination point and a single-pole double-throw switch, which can simultaneously train the traveling wave characteristics of the high-power and low-power regions by a large order of magnitude.

[0176] In one embodiment, in order to improve the accuracy of the large-signal modeling model of the GaN microwave device, after constructing the large-signal behavior model according to the voltage-current characteristics, the following steps can be further performed:

[0177] First, receive the verification traveling wave characteristics. The verification traveling wave characteristics include the actual data of the verification traveling wave characteristics. Specifically, in order to verify whether the model is accurate, new traveling wave characteristics, that is, the verification traveling wave characteristics, need to be received for verification, and the verification traveling wave characteristics include the actual data of the verification traveling wave characteristics.

[0178] Then, use the large-signal behavior model to simulate and verify the traveling wave characteristics to obtain simulation data. Specifically, as Figure 13 shown, the figure shows the output power, gain, and efficiency of the DNN model during verification. The large-signal behavior model simulates according to the verification traveling wave characteristics to obtain simulation data.

[0179] Next, determine whether the error between the actual data and the simulation data is within the allowable range. If the error is within the allowable range, it is determined that the large-signal behavior model can be put into use. If the error is not within the allowable range, it is determined that the large-signal behavior model cannot be put into use. Specifically, determining whether the error between the actual data and the simulation data is within the allowable range is to verify the simulation accuracy of the large-signal behavior model. If the error is within the allowable range, it means that the large-signal behavior model has high accuracy and can be put into use; otherwise, it means that the large-signal behavior model has low accuracy and cannot be put into use.

[0180] Perform simulation comparison on the obtained large-signal behavior model, verify the function of the large-signal behavior model, and ensure the accuracy of the large-signal behavior model.

[0181] In addition, it should be noted that when performing large-signal tests on GaN transistors, traveling wave characteristics, frequency, and voltage will be obtained. Through simulation training at different frequencies, the fitting accuracy within a wide dynamic range can be improved. As Figure 14 shown, the figure shows the output power, gain, and efficiency of the DNN model. It can be seen that the large-signal model can not only perform modeling at the transistor level, but also perform accurate modeling at the integrated circuit module level to complete system design.

[0182] Based on the above method, an embodiment of the present application further discloses a large-signal model modeling system based on DNN. The system is based on a GaN microwave device including a GaN transistor. As Figure 15 , the system includes the following modules:

[0183] A traveling-wave characteristic acquisition module 1501, configured to perform a large-signal test on the GaN transistor to obtain traveling-wave characteristics, where the traveling-wave characteristics include input-layer characteristics and output-layer characteristics;

[0184] A traveling-wave characteristic processing module 1502, configured to process the input-layer characteristics to obtain input-layer data, and process the output-layer characteristics to obtain output-layer data;

[0185] A high-power region model training module 1503, configured to extract the real-part data and imaginary-part data of the input-layer data and the output-layer data, and use DNN to train a high-power region deep neural network model according to the real-part data and the imaginary-part data;

[0186] A low-power region model training module 1504, configured to extract the amplitude data and phase data of the input-layer data and the output-layer data, and use DNN to train a low-power region deep neural network model according to the amplitude data and the phase data;

[0187] A voltage-current characteristic conversion module 1505, configured to train the traveling-wave characteristics according to the high-power region deep neural network model and the low-power region deep neural network model to obtain training traveling-wave characteristics, and convert the training traveling-wave characteristics into voltage-current characteristics;

[0188] A behavior model construction module 1506, configured to construct a large-signal behavior model according to the voltage-current characteristics.

[0189] In one embodiment, the traveling-wave characteristic acquisition module 1501 is specifically configured to obtain the fundamental wave condition and harmonic condition for testing; perform an input power scan on the GaN transistor through a vector network analyzer under the fundamental wave condition and the harmonic condition to obtain a scan result; and read the scan result through a coupler to obtain the traveling-wave characteristics.

[0190] In one embodiment, the traveling-wave characteristic processing module 1502 is specifically configured to calculate the target phase of the fundamental wave incident wave at the input port; normalize the remaining incident waves according to the target phase to obtain the input-layer data; and normalize the reflected waves according to the target phase to obtain the output-layer data.

[0191] In one embodiment, the high-power region model training module 1503 is specifically configured to extract the target amplitude of the fundamental incident wave at the input port, extract the real part data and the imaginary part data of the input layer data, and extract the real part data and the imaginary part data of the output layer data; normalize the target amplitude, the real part data and the imaginary part data of the input layer data, and the real part data and the imaginary part data of the output layer data to obtain high-power region training data; import the high-power region training data into a preset first neural network for training and calculate a first loss value; determine whether the first loss value is less than a preset value; if the first loss value is less than the preset value, stop training and obtain a high-power region deep neural network model according to the trained first neural network.

[0192] In one embodiment, the low-power region model training module 1504 is specifically configured to extract the dB amplitude of the fundamental incident wave at the input port, where the unit of the dB amplitude is dB, extract the amplitude data and the phase data of the input layer data, and extract the amplitude data and the phase data of the output layer data; normalize the dB amplitude, the amplitude data and the phase data of the input layer data, and the amplitude data and the phase data of the output layer data to obtain low-power region training data; import the low-power region training data into a preset second neural network for training and calculate a second loss value; determine whether the second loss value is less than a second preset value; if the second loss value is less than the second preset value, stop training and obtain a low-power region deep neural network model according to the trained second neural network.

[0193] In one embodiment, the voltage-current characteristic conversion module 1505 is specifically configured to set a determination point for the high and low power regions; construct a partitioned deep neural network model according to the determination point, the high-power region deep neural network model, and the low-power region deep neural network model; train the traveling wave characteristics according to the partitioned deep neural network model and perform anti-normalization processing to obtain the trained traveling wave characteristics.

[0194] An embodiment of the present application also discloses a computer device.

[0195] Specifically, the computer device includes a memory and a processor, and a computer program capable of being loaded and executed by the processor is stored on the memory.

[0196] An embodiment of the present application also discloses a computer-readable storage medium.

[0197] Specifically, the computer-readable storage medium stores a computer program capable of being loaded and executed as described above. The computer-readable storage medium includes, for example, various media that can store program codes, such as a USB flash drive, a portable hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc.

[0198] This specific embodiment is only an interpretation of the present invention and not a limitation thereof. After reading this specification, those skilled in the art may make modifications to this embodiment that do not contribute creatively, but as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. A large signal modeling method based on DNN, the method is based on a GaN microwave device including a GaN transistor, characterized in that: The method comprises: Performing a large signal test on the GaN transistor to obtain traveling wave characteristics, wherein the traveling wave characteristics include input layer characteristics and output layer characteristics; Processing the input layer characteristics to obtain input layer data, and processing the output layer characteristics to obtain output layer data; Extracting the real data and the imaginary data of the input layer data and the output layer data, and obtaining a high-power area deep neural network model by using DNN training according to the real data and the imaginary data; Extracting amplitude data and phase data of the input layer data and the output layer data, and obtaining a low-power area deep neural network model by using DNN training according to the amplitude data and the phase data; Training the traveling wave characteristics according to the high-power area deep neural network model and the low-power area deep neural network model to obtain trained traveling wave characteristics, and converting the trained traveling wave characteristics into voltage and current characteristics; A large signal behavior model is constructed according to the voltage and current characteristics.

2. The method according to claim 1, characterized in that: The GaN transistor is subjected to large signal testing to obtain traveling wave characteristics including: Obtain fundamental wave conditions and harmonic conditions for testing; Under the fundamental wave condition and the harmonic wave condition, an input power scan of the GaN transistor is performed by a vector network analyzer to obtain a scan result; The scanning result is read through a coupler to obtain the traveling wave characteristic.

3. The method according to claim 1, characterized in that: The input layer characteristic includes an incident wave, the output layer characteristic includes a reflected wave, and the processing of the input layer characteristic to obtain input layer data and the processing of the output layer characteristic to obtain output layer data include: Calculate the target phase of the fundamental incident wave at the input port; Normalizing the remaining incident waves according to the target phase to obtain input layer data; The reflected wave is normalized according to the target phase to obtain output layer data.

4. The method according to claim 3, characterized in that: The extracting the real data and the imaginary data of the input layer data and the output layer data comprises: Extracting the target amplitude of the fundamental wave incident wave of the input port, extracting the real data and the imaginary data of the input layer data, and extracting the real data and the imaginary data of the output layer data; The obtaining of a high-power area deep neural network model by using DNN training according to the real data and the imaginary data comprises: Normalizing the target amplitude, the real data and the imaginary data of the input layer data, and the real data and the imaginary data of the output layer data to obtain high-power area training data; Importing the high-power area training data into a preset first neural network for training and calculating a first loss value; Determining whether the first loss value is less than a preset value; If the first loss value is less than a preset value, the training is stopped and a high-power area deep neural network model is obtained based on the trained first neural network.

5. The method according to claim 4, characterized in that: The extracting the amplitude data and the phase data of the input layer data and the output layer data comprises: Extract the dB amplitude of the fundamental wave incident wave of the input port, the unit of the dB amplitude is dB, extract the amplitude data and phase data of the input layer data, and extract the amplitude data and phase data of the output layer data; The obtaining of a low-power area deep neural network model by using DNN training according to the amplitude data and the phase data comprises: Normalizing the dB amplitude, the amplitude data and the phase data of the input layer data, and the amplitude data and the phase data of the output layer data to obtain low-power area training data; Importing the low-power area training data into a preset second neural network for training and calculating a second loss value; Determining whether the second loss value is less than a second preset value; If the second loss value is less than a second preset value, the training is stopped and a low-power area deep neural network model is obtained based on the trained second neural network.

6. The method according to claim 1, characterized in that: The step of training the travel wave characteristic according to the high power area deep neural network model and the low power area deep neural network model to obtain the trained travel wave characteristic comprises: Set the decision points for high and low power areas; Constructing a partitioned deep neural network model according to the decision point, the high-power area deep neural network model, and the low-power area deep neural network model; The travel wave characteristics are trained according to the partitioned deep neural network model and denormalized to obtain the trained travel wave characteristics.

7. The method according to claim 1, characterized in that: After the large signal behavior model is constructed according to the voltage and current characteristics, the method further includes: Receiving verification of the travel wave characteristic, the verification of the travel wave characteristic comprising actual data of the verification of the travel wave characteristic; The verification wave characteristics are simulated using the large signal behavior model to obtain simulation data: Determine whether the error between the actual data and the simulation data is within an allowable range; If the error is within the allowable range, it is determined that the large signal behavior model can be put into use; If the error is not within the allowable range, it is determined that the large signal behavior model cannot be put into use.

8. A large signal modeling system based on DNN, characterized in that: The system is based on a GaN microwave device including a GaN transistor, and the system comprises: A traveling wave characteristic acquisition module (1501) is used to perform a large signal test on the GaN transistor to obtain traveling wave characteristics, wherein the traveling wave characteristics include input layer characteristics and output layer characteristics; A wave characteristic processing module (1502), used for processing the input layer characteristic to obtain input layer data, and processing the output layer characteristic to obtain output layer data; A high-power area model training module (1503) is used to extract the real data and the imaginary data of the input layer data and the output layer data, and obtain a high-power area deep neural network model by using DNN training according to the real data and the imaginary data; A low power area model training module (1504), used to extract the amplitude data and phase data of the input layer data and the output layer data, and obtain a low power area deep neural network model by using DNN training according to the amplitude data and the phase data; A voltage-current characteristic conversion module (1505) is used to train the traveling wave characteristic according to the high-power area deep neural network model and the low-power area deep neural network model to obtain a training traveling wave characteristic, and convert the training traveling wave characteristic into a voltage-current characteristic; The behavior model construction module (1506) is used to construct a large signal behavior model according to the voltage and current characteristics.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program which can be loaded by the processor and executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored which can be loaded by a processor and execute the method according to any one of claims 1 to 7.

Images