CN119720906A - A method, system, device and medium for optimizing analog equalizer design based on deep learning algorithm - Google Patents
A method, system, device and medium for optimizing analog equalizer design based on deep learning algorithm Download PDFInfo
- Publication number
- CN119720906A CN119720906A CN202411795184.5A CN202411795184A CN119720906A CN 119720906 A CN119720906 A CN 119720906A CN 202411795184 A CN202411795184 A CN 202411795184A CN 119720906 A CN119720906 A CN 119720906A
- Authority
- CN
- China
- Prior art keywords
- neural network
- equalizer circuit
- equalizer
- trained
- predicted
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Filters That Use Time-Delay Elements (AREA)
- Cable Transmission Systems, Equalization Of Radio And Reduction Of Echo (AREA)
- Feedback Control In General (AREA)
Abstract
一种基于深度学习算法的模拟均衡器优化设计方法、系统、设备及介质，该方法首先通过均衡器电路电磁响应曲线对逆向神经网络进行训练，通过训练好的逆向神经网络生成额外的均衡器电路参数数据，并与原始仿真数据组合成新的数据集，以达到解决仿真软件耗时过长的问题，通过此新的数据集对正向神经网络进行训练，训练好的正向神经网络输出预测的均衡器电磁响应曲线，将其与目标均衡器电磁响应曲线相比较，符合条件即利用逆向神经网络生成实际符合条件的均衡器电路参数，并完成设计；系统、设备及介质用于承载和实现所述方法；本发明通过深度学习方法优化设计流程，实现了电路设计的自动化，减少设计时间和计算复杂度。
A method, system, device and medium for optimizing the design of an analog equalizer based on a deep learning algorithm. The method first trains an inverse neural network through an electromagnetic response curve of an equalizer circuit, generates additional equalizer circuit parameter data through the trained inverse neural network, and combines the data with the original simulation data into a new data set to solve the problem that the simulation software takes too long. The forward neural network is trained through the new data set, and the trained forward neural network outputs a predicted equalizer electromagnetic response curve, which is compared with a target equalizer electromagnetic response curve. If the conditions are met, the inverse neural network is used to generate actual equalizer circuit parameters that meet the conditions, and the design is completed. The system, device and medium are used to carry and implement the method. The present invention optimizes the design process through a deep learning method, realizes the automation of circuit design, and reduces design time and computational complexity.
Description
Technical Field
The invention belongs to the technical field of analog equalizer, and particularly relates to an analog equalizer optimal design method, system, equipment and medium based on a deep learning algorithm.
Background
In the fields of wireless communication and precise measurement, system performance is often limited by non-ideal factors such as amplitude-frequency response and group delay characteristics of links. The effects of these non-ideal characteristics are particularly pronounced when the system is operating in a wide frequency band, which in turn has a serious negative impact on the integrity and accuracy of the signal transmission. In the field of precision measurement, solid-state power amplifiers are used as core components, whose design criteria play a decisive role in overall performance. However, due to the inherent nonlinear characteristics of the devices or the problems of multistage cascade connection of the power circuits, signal distortion is easy to be induced due to the uneven characteristics of transmission parameters of the two-port and multi-port devices, so that the output amplitude of the system in a wide frequency band range has large fluctuation, the waveform of the amplifying circuit is deteriorated, intersymbol interference is caused, and finally the communication quality of the system is caused.
To overcome these effects, a strategy that is currently commonly adopted in the industry is to introduce equalizers at key nodes. The equalizer has the main functions of compensating the non-uniformity and the non-flatness caused by elements and channels, reducing signal distortion and improving the stability of output power. Through the optimization, the performance of the system in broadband application can be obviously improved, and the definition and reliability of signals are enhanced.
However, with the development of communication technology, the requirement for broadband and high gain characteristics of the power device is increasing, which results in that more target parameters and design standards need to be considered in the design process, and the design difficulty of the equalizer is increased. Particularly in the design of microwave power amplifiers, when the device structure becomes highly complex, the equalizer design and optimization process may become extremely time consuming, and using conventional design methods, it may take hours to days to complete an equalizer design. In recent years, as artificial intelligence is applied in more and more fields, the neural network has good robustness and generalization capability, and is suitable for rapid design under different indexes. Therefore, the neural network algorithm is introduced in the design process, so that the multi-target and multi-index design can be conveniently carried out.
The existing implementation scheme is as follows:
1. in the traditional circuit design process, a proper equalizer circuit structure is selected through a specified target index, then a system function expression of the circuit is calculated manually according to the circuit structure, and related parameter values of the circuit are calculated through related requirements.
2. In the process of designing the neural network by microwaves, the design of the filter is usually carried out by the neural network at present, the method is limited to the filter with a specific frequency band, the structure of the filter is required to be specified, a great amount of training data and computing resources are required to obtain a good effect, and a model can be required to be retrained for different application scenes.
3. In the existing neural network training mode, the neural network is trapped in a local optimal solution due to the influence of a non-unique solution due to the fact that the data sets are too few, the time consumption is too long due to the fact that the data sets are obtained through simulation software, the requirement on an initial value is high due to the fact that optimization algorithms based on particle swarm and the like are high, the simulation software is required to be used continuously in the iteration process, and time is easy to consume.
The disadvantages of the prior art mainly include:
1. High cost and resource consumption traditional analog equalizer design methods rely on complex electromagnetic simulations and difficult computational designs, which not only increase design costs, but also require significant computational resources.
2. Inefficiency and long period-the manual design and adjustment process is time consuming and inefficient, particularly when faced with complex circuit designs, requiring a large number of iterations and trials to be performed manually, which lengthens the design period.
3. Design complexity as communication systems demand high performance channel characteristics, equalizer designs become more complex and conventional design methods have difficulty coping with ever-increasing design criteria and target parameters.
4. Depending on expertise, existing design methods rely on the expertise and experience of the designer, and are difficult for the general population to quickly master and apply to actual designs.
The patent application document with publication number CN 113128119A discloses a filter reverse design and optimization method based on deep learning, wherein structural parameters of the filter are designed through electromagnetic response change of the filter, forward network generation data are utilized to train a reverse network, an electromagnetic response curve of the filter is obtained through Chebyshev polynomial synthesis, and finally the optimized filter response curve and the structural parameters of the filter are output. However, according to the method, the electromagnetic response of the filter is obtained through the Chebyshev polynomial synthesis, the Chebyshev polynomial with proper order is required to be selected manually, and the error caused by manual selection is larger when the order is high, so that the generated precision is reduced, and the calculation cost is wasted.
Disclosure of Invention
In order to overcome the defects in the prior art, the invention aims to provide a simulation equalizer optimal design method, a system, equipment and a medium based on a deep learning algorithm, wherein the simulation equalizer circuit is designed and optimized by the deep learning method, additional equalizer circuit parameter data is self-generated by using a reverse neural network and forms a new data set with original simulation data for training of a forward neural network, and finally equalizer circuit parameters predicted by a reverse neural network model and an equalizer electromagnetic response curve predicted by the forward neural network are obtained, so that the automatic design and optimization of the simulation equalizer circuit are realized, and the design efficiency and the accuracy are improved.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
a simulation equalizer optimization design method based on a deep learning algorithm comprises the following steps:
step 1, taking an electromagnetic response curve of a randomly generated equalizer circuit as an input of a reverse neural network, training the reverse neural network to obtain a trained reverse neural network, and generating additional equalizer circuit parameter data by using the trained reverse neural network;
Step 2, combining the additional equalizer circuit parameter data generated by the trained reverse neural network in the step 1 with the original simulation data to form a new data set, wherein the new data set is used for training the forward neural network, and after the convergence condition is reached, the trained forward neural network is obtained;
And step 3, inputting the expected equalizer circuit electromagnetic response curve into the reverse neural network trained in the step 1, generating predicted equalizer circuit parameter data through the reverse neural network, inputting the generated predicted equalizer circuit parameter data into the forward neural network trained in the step 2, outputting a predicted actual equalizer circuit electromagnetic response curve through the forward neural network, inputting the predicted actual equalizer circuit electromagnetic response curve into the reverse neural network trained in the step 1 to obtain actual equalizer circuit parameter data, and finally obtaining the actual equalizer circuit parameter data predicted through the reverse neural network and the actual equalizer circuit electromagnetic response curve predicted by the forward neural network.
The reverse neural network comprises an input layer, four fully-connected hidden layers, a Dropout layer, a LeakyReLU activation function and an output layer, wherein the number of neurons of the four fully-connected hidden layers is 251, 128, 64 and 10 respectively, and the training target of the reverse neural network is that the predicted equalizer circuit parameter data and the loss function of the real equalizer circuit parameter are minimum.
The forward neural network comprises an input layer, a fully-connected hidden layer containing LeakyReLU activation functions, a hidden layer containing Dropout layers and LeakyReLU activation functions and an output layer, wherein the number of neurons of the input layer is set to 11, the number of neurons of the hidden layer is set to 80, and the training target of the forward neural network is the predicted return loss of an electromagnetic response curve, the return loss of the insertion loss curve of the electromagnetic response curve corresponding to the equalizer circuit parameters in the original data set and the loss function of the insertion loss curve are minimum.
The invention also provides an analog equalizer optimization design system based on the deep learning algorithm, which comprises:
The additional equalizer circuit parameter data generation module is used for taking the electromagnetic response curve of the equalizer circuit which is randomly generated as the input of the reverse neural network, training the reverse neural network to obtain a trained reverse neural network, and generating additional equalizer circuit parameter data by using the trained reverse neural network;
The forward neural network training module is used for combining the additional equalizer circuit parameter data generated by the trained reverse neural network with the original simulation data to form a new data set, wherein the new data set is used for training the forward neural network, and the trained forward neural network is obtained after the convergence condition is reached;
The system comprises an actual equalizer circuit parameter data and electromagnetic response curve generation module, an actual equalizer circuit parameter data obtaining module and an actual equalizer circuit parameter data obtaining module, wherein the actual equalizer circuit parameter data and the electromagnetic response curve generation module are used for inputting an expected equalizer circuit electromagnetic response curve into a trained reverse neural network, generating predicted equalizer circuit parameter data through the reverse neural network, inputting the generated predicted equalizer circuit parameter data into a trained forward neural network, outputting the predicted actual equalizer circuit electromagnetic response curve through the forward neural network, finally inputting the predicted actual equalizer circuit electromagnetic response curve into the trained reverse neural network in the step 1 to obtain the actual equalizer circuit parameter data, and finally obtaining the actual equalizer circuit parameter data predicted through the reverse neural network and the actual equalizer circuit electromagnetic response curve predicted through the forward neural network.
The invention also provides an analog equalizer optimal design device based on the deep learning algorithm, which comprises:
The memory is used for storing the computer program of the simulation equalizer optimal design method based on the deep learning algorithm and is equipment readable by a computer;
and the processor is used for realizing the simulated equalizer optimal design method based on the deep learning algorithm when executing the computer program.
The invention also provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program can realize the simulation equalizer optimization design method based on the deep learning algorithm when being executed by a processor.
Compared with the prior art, the invention has the beneficial effects that:
1. and the manual calculation cost is reduced.
According to the invention, a design mode of combining the forward neural network and the reverse neural network is adopted, partial extra circuit parameter data generated by the trained reverse neural network and the original simulation data are combined into a new data set, the forward neural network is trained by the new data set, the trained reverse neural network can generate equalizer circuit parameters meeting the design requirements, the trained forward neural network can generate an equalizer electromagnetic response curve meeting the design requirements, so that any complex equalizer circuit design can be realized, the design of any equalizing effect can be basically realized within the frequency range of 0-10GHz, the accurate design result can be obtained within a short time, and the manual complex calculation and multiple iterations are avoided.
2. Saving computing resources and design time
According to the invention, the forward neural network is adopted to predict the electromagnetic response curve of the equalizer, the reverse neural network is adopted to predict the equalizer circuit parameters, only a small amount of random equalizer circuit parameters and the simulated electromagnetic response curve are needed to be used as input data, a large amount of new equalizer circuit electromagnetic simulation results can be automatically generated through the reverse neural network, and the time cost and calculation resources for repeatedly calling the electromagnetic simulation software are saved.
3. And the design efficiency is improved.
According to the design method combining the reverse neural network and the forward neural network, the reverse neural network can automatically generate a large amount of data for training the forward neural network by using a small amount of initial data sets, so that the data set scale is increased, a large amount of time consumption caused by repeatedly calling simulation software is avoided, and the model training time and the design period are greatly reduced. And the trained forward neural network has strong response speed, and for 1 design, only 1.35s is required, and for 1000 random different designs, only 2.67s is required. And when aiming at a large number of equalizers with different design indexes, the equalizer design result can be output in a very short time, so that the design efficiency is greatly improved.
4. Reducing design complexity
The invention can meet the design requirement of complex equalizer with high performance, wide frequency band and multiple design indexes by automatically designing complex equalizer circuit through deep learning model, so that the circuit design can be accurately and efficiently completed when the design standard of increasingly complex communication system is faced.
5. Simplifying the design difficulty
According to the invention, the design result is automatically generated through the deep learning algorithm, so that the traditional calculation and optimization flow is simplified. Through the trained neural network, a user can design the equalizer circuit meeting the requirements only by providing corresponding design indexes, so that a non-professional person can quickly get on hand, the efficient design of the equalizer circuit is realized, and the dependence on professional knowledge is reduced.
6. Optimizing design accuracy
The invention can realize the design of any complex equalization circuit by utilizing the excellent fitting effect of the neural network on the nonlinear function, and the matching effect of the return loss and the insertion loss curve of the corresponding electromagnetic response curve and the actual result is good, and the design error is less than 1dB.
7. Has prospective effect
The invention designs the passive equalizer by utilizing deep learning, can explore a wider design parameter space by the deep learning, and discovers an innovative design scheme which is possibly ignored by the traditional design method. The invention can meet the requirements of future experimental scenes on various equalizers on complexity and severity as a prospective technical scheme.
In summary, the equalizer circuit is designed by the deep learning method, and a method of combining the reverse neural network and the forward neural network is adopted to automatically generate a large amount of data for training, so that the manpower cost and the time consumption are saved, and the method has the advantages of reducing the manual calculation cost and improving the design efficiency and the design precision under the condition that the design precision meets the requirements.
Drawings
Fig. 1 is an overall block diagram of an equalizer design method of the present invention.
Fig. 2 is a circuit configuration diagram of an equalizer of the present invention.
Fig. 3 is a diagram of the inverse neural network of the present invention.
Fig. 4 is a block diagram of a forward neural network of the present invention.
Fig. 5 is a graph showing the prediction result of the return loss curve according to the present invention.
Fig. 6 is a graph of the insertion loss curve prediction result of the present invention.
Detailed Description
The invention is further described below with reference to the accompanying drawings and specific examples.
Aiming at the common problems of the traditional analog passive equalizer circuit comprehensive design method, especially when the application scene of the method gradually shows that the working frequency band is wider and wider, the required compensation gain is larger and the design requirements such as the application index requirement are larger and larger, the traditional equalizer circuit comprehensive design method depends on manual calculation and continuous adjustment, the method for calculating the transfer function and then selecting the circuit parameters is time-consuming and complex in calculation process, and the circuit design method depends on manual calculation is more difficult when the design index is gradually complex.
In order to solve the problem, the invention provides an innovative method, and the equalizer design of various complex indexes can be easily realized by using a two-layer neural network. According to the method, firstly, an equalizer circuit electromagnetic response curve is used for training a reverse neural network, additional equalizer circuit parameter data is generated through the trained reverse neural network, and the additional equalizer circuit parameter data and original simulation data are combined into a new data set, so that the problem that simulation software consumes too much time is solved, the forward neural network is trained through the new data set, the trained forward neural network outputs a predicted equalizer electromagnetic response curve, the predicted equalizer electromagnetic response curve is compared with a target equalizer electromagnetic response curve, and the equalizer circuit parameter which meets the actual conditions is generated through the reverse neural network, so that the design is completed. According to the invention, the design flow is optimized through the deep learning method, so that the automation of circuit design is realized, and the design time and the calculation complexity are reduced.
The invention solves the following technical problems:
1. In order to solve the problems of high cost and high workload of the traditional design method depending on manual calculation and continuous optimization adjustment, the invention needs to realize automatic design of the circuit needing equalization, including selection of a proper equalization circuit structure.
2. When a large number of equalizers are designed, a large amount of time and calculation resources are needed by the traditional method, and the invention can realize the design of a large number of circuits in a very short time and realize the design with high efficiency and high precision.
The following describes the design of a T-equalizer circuit in more detail.
As shown in fig. 1, the method for optimizing the design of the analog equalizer based on the deep learning algorithm comprises the following steps:
step 1, generating additional equalizer circuit parameter data by using a reverse neural network;
The electromagnetic response curve of 1000 groups of equalizer circuits generated randomly is used as the input of the reverse neural network to train the reverse neural network, the output of the reverse neural network is equalizer circuit parameter data, the training target is that the predicted equalizer circuit parameter data and the loss function of the real equalizer circuit parameters are minimum, after the convergence condition or the maximum iteration number is reached, the trained reverse neural network is obtained, the trained reverse neural network is used for generating additional equalizer circuit parameter data, the reverse neural network can automatically generate more equalizer circuit parameter data, and a great amount of time consumed by calling simulation software to generate data is compensated.
Fig. 2 is a schematic diagram of an equalizer circuit structure used in the present invention, wherein a lumped parameter circuit adopts a broadband network model with a combination of broadband series and parallel.
The ten key circuit parameters of the equalizer are C1, C2, C3, L1, L2, L3, R1, R2, R3 and R4, and proper parameter ranges are set, 1000 return loss and insertion loss curves are generated by electromagnetic simulation software in the ranges, the frequency of the electromagnetic response curve is 3-10GHz, and the return loss is required to be smaller than-10 dB at 3-10 GHz. There are 251 data points for each curve over this frequency range. The 1000 sets of samples were made into a dataset, 900 sets were training sets, 100 sets were test sets, and the inverse neural network of fig. 3 was trained. The trained reverse neural network automatically generates 3600 groups of data, and forms 4500 groups of data together with the previous 900 groups of arrays.
As shown in fig. 3, the inverse neural network includes an input layer, four fully-connected hidden layers, and corresponding Dropout layers, leakyReLU activation functions, and an output layer, where the number of neurons of the four fully-connected hidden layers is 251, 128, 64, and 10, the Dropout values are all 0.07, the loss function is an MSR mean square error function, the maximum number of iterations is 1000, and the training objective is that the predicted equalizer circuit parameter data and the loss function of the real equalizer circuit parameter are minimum.
Step 2, combining the extra equalizer circuit parameter data generated by the trained reverse neural network in the step 1 with the original simulation data to form a new data set, wherein the new data set is used for training the forward neural network, and after the convergence condition is reached, the trained forward neural network is obtained, and the forward neural network is used for generating a predicted equalizer circuit electromagnetic response curve;
The 4500 groups of data generated in the step 1 are used for training a forward neural network, and the forward neural network can obtain return loss and insertion loss of corresponding frequencies by providing predicted frequency information. Repeating 251 times, and combining to obtain the return loss and insertion loss curve predicted by the forward neural network in the frequency range of 3-10 GHz.
As shown in fig. 4, the forward neural network includes an input layer, a fully connected hidden layer including LeakyReLU activation functions, a hidden layer including Dropout layer and LeakyReLU activation functions, and an output layer, wherein the input layer has 11 neurons, the hidden layer has 80 neurons, the loss function is an MSR mean square error function, the maximum iteration number is 1000, and the training target is the predicted return loss of the electromagnetic response curve, the return loss of the electromagnetic response curve corresponding to the equalizer circuit parameters in the original dataset, and the loss function of the insertion loss curve is minimum.
And 3, inputting the expected equalizer circuit electromagnetic response curve into the reverse neural network trained in the step 1 after the forward neural network is trained, generating predicted equalizer circuit parameter data through the reverse neural network, inputting the generated predicted equalizer circuit parameter data into the forward neural network trained in the step 2, outputting a predicted actual equalizer circuit electromagnetic response curve through the forward neural network, comparing the predicted actual equalizer circuit electromagnetic response curve with the expected equalizer circuit electromagnetic response curve, if the error is larger, optimizing the forward neural network, optimizing and stopping when the loss function reaches a threshold value or the maximum iteration step number, finally inputting the predicted actual equalizer circuit electromagnetic response curve into the reverse neural network trained in the step 1, and finally obtaining the predicted actual equalizer circuit parameter data through the reverse neural network and the predicted actual equalizer circuit electromagnetic response curve of the forward neural network.
And (3) taking the error of the predicted equalizer circuit electromagnetic response curve generated in the step (2) and the required target equalizer electromagnetic response curve as an optimization target, wherein the loss functions of the forward neural network and the reverse neural network are MSE (mean square error) functions, and when the loss functions reach a threshold value or the maximum iteration step number, the optimization is stopped, wherein the output of the reverse neural network in the step (1) is a circuit parameter of an equalizer, and the output of the forward neural network in the step (2) is an electromagnetic response curve of the equalizer circuit.
Fig. 5 and 6 show the equalizer design results. The horizontal axis of FIG. 5 is frequency and the vertical axis is dB (S11), the design goal is to make dB (S11) < -10dB in 5GHz-10GHz, and the target curve is shown by the solid line in the figure. The dashed line in the figure is the result of the equalizer circuit parameter result predicted by the inverse neural network in simulation software, the horizontal axis in fig. 6 is the frequency, the vertical axis is dB (S21), the goal is to make dB (S21) appear monotonously increasing in 5GHz-10GHz, and approximately 0dB at 10GHz, and the target curve is shown by the solid line in the figure. The dashed line in the figure is the result of the equalizer circuit parameter result predicted by the inverse neural network in the simulation software.
The horizontal axis of FIG. 5 is frequency and the vertical axis is dB (S11), the goal is to make dB (S11) < -10dB in 5GHz-10GHz, and the target curve is shown by the solid line in the figure. The dashed line in the figure is the result of the circuit parameter result predicted by the inverse neural network in the simulation software.
The horizontal axis of FIG. 6 is frequency and the vertical axis is dB (S21), the goal is to make dB (S21) appear monotonically increasing in 5GHz-10GHz, and approximately 0dB at 10GHz, the target curve is shown by the solid line in the figure. The dashed line in the figure is the result of the circuit parameter result predicted by the inverse neural network in the simulation software.
From fig. 5 and 6, it can be seen that the equalizer designed by the invention has a good match with the target curve, the average absolute error of the return loss curve is 0.9665dB, the average absolute error of the insertion loss curve is 0.2447dB, and the design errors are all less than 1dB, so as to meet the design requirement. Therefore, the trained neural network can be used for designing any complex equalizer circuit by only providing corresponding related indexes, such as the equalizer designed in the range of 5-10GHz and only providing the change condition of two electromagnetic response curves in the frequency range.
The key points and the protection points of the invention are as follows:
1. The invention adopts a forward and reverse two-layer neural network to design, realizes the automatic design of the simulated equalizer, and remarkably improves the design efficiency and precision.
2. The reverse neural network comprises an input layer, four fully-connected hidden layers, a corresponding Dropout layer, leakyReLU activation functions and an output layer, wherein the number of neurons of the four fully-connected hidden layers is 251, 128, 64 and 10, the value of Dropout is 0.07, the loss function is an MSR mean square error function, the maximum iteration number is 1000, and the training target is that the loss function of predicted equalizer circuit parameter data and real equalizer circuit parameters is minimum. Judging whether the network is trained well by examining the convergence of the loss functions of the training set and the testing set and the predicted performance of the network on the new data set, and if the trained network has poor performance, adjusting the super parameters such as the size of the data set, the training period, the loss function and the like.
3. The forward neural network comprises an input layer, a fully-connected hidden layer containing LeakyReLU activation functions, a hidden layer containing Dropout layers and LeakyReLU activation functions and an output layer, wherein the input layer is provided with 11 neurons, the hidden layer contains 80 neurons, the loss function is an MSR mean square error function, the maximum iteration number is 1000, and the training target is the predicted return loss of an electromagnetic response curve, the return loss of an electromagnetic response curve of an insertion loss curve corresponding to equalizer circuit parameters in an original dataset and the loss function of the insertion loss curve are minimum. Judging whether the network is trained well by examining the convergence of the loss functions of the training set and the testing set and the predicted performance of the network on the new data set, and if the trained network is poor in performance, adjusting the super parameters such as the size of the data set, the training period, the loss functions and the like.
The invention provides an analog equalizer optimal design method based on a deep learning algorithm, which aims to solve the problems of time consumption, low efficiency and high cost of the traditional design method in complex circuit design. The invention realizes the automatic design of the analog equalizer by utilizing the two layers of neural networks to design the analog equalizer through deep learning.
The invention also provides an analog equalizer optimization design system based on the deep learning algorithm, which comprises:
The additional equalizer circuit parameter data generation module is used for taking the electromagnetic response curve of the equalizer circuit generated randomly in the step 1 as the input of the reverse neural network to train the reverse neural network, obtaining a trained reverse neural network and generating additional equalizer circuit parameter data by using the trained reverse neural network;
The forward neural network training module is used for combining the additional equalizer circuit parameter data generated by the reverse neural network trained in the step 2 with the original simulation data to form a new data set, wherein the new data set is used for training the forward neural network, and the trained forward neural network is obtained after the convergence condition is reached;
The actual equalizer circuit parameter data and electromagnetic response curve generating module is used for inputting the expected equalizer circuit electromagnetic response curve in the step 3 into the reverse neural network trained in the step 1, generating predicted equalizer circuit parameter data through the reverse neural network, inputting the generated predicted equalizer circuit parameter data into the forward neural network trained in the step 2, outputting the predicted actual equalizer circuit electromagnetic response curve through the forward neural network, finally inputting the predicted actual equalizer circuit electromagnetic response curve into the reverse neural network trained in the step 1 to obtain actual equalizer circuit parameter data, and finally obtaining the actual equalizer circuit parameter data predicted through the reverse neural network and the actual equalizer circuit electromagnetic response curve predicted by the forward neural network.
The invention also provides an analog equalizer optimal design device based on the deep learning algorithm, which comprises:
The memory is used for storing the computer program of the simulation equalizer optimal design method based on the deep learning algorithm and is equipment readable by a computer;
and the processor is used for realizing the simulated equalizer optimal design method based on the deep learning algorithm when executing the computer program.
The invention also provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program can realize the simulation equalizer optimization design method based on the deep learning algorithm when being executed by a processor.
Claims (6)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202411795184.5A CN119720906B (en) | 2024-12-09 | 2024-12-09 | A method, system, device and medium for optimizing analog equalizer design based on deep learning algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202411795184.5A CN119720906B (en) | 2024-12-09 | 2024-12-09 | A method, system, device and medium for optimizing analog equalizer design based on deep learning algorithm |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN119720906A true CN119720906A (en) | 2025-03-28 |
| CN119720906B CN119720906B (en) | 2025-08-26 |
Family
ID=95097947
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202411795184.5A Active CN119720906B (en) | 2024-12-09 | 2024-12-09 | A method, system, device and medium for optimizing analog equalizer design based on deep learning algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN119720906B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| FR706006A (en) * | 1929-10-30 | 1931-06-17 | Materiel Telephonique | Improvements to electrical devices referred to as balancing or compensating networks |
| WO2021007812A1 (en) * | 2019-07-17 | 2021-01-21 | 深圳大学 | Deep neural network hyperparameter optimization method, electronic device and storage medium |
| CN113128119A (en) * | 2021-04-21 | 2021-07-16 | 复旦大学 | Filter reverse design and optimization method based on deep learning |
| WO2023038070A1 (en) * | 2021-09-10 | 2023-03-16 | Mitsubishi Electric Corporation | Multiplier-less sparse deep neural network |
| US20230081531A1 (en) * | 2021-09-10 | 2023-03-16 | Mitsubishi Electric Research Laboratories, Inc. | Multiplier-Less Sparse Deep Neural Network |
-
2024
- 2024-12-09 CN CN202411795184.5A patent/CN119720906B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| FR706006A (en) * | 1929-10-30 | 1931-06-17 | Materiel Telephonique | Improvements to electrical devices referred to as balancing or compensating networks |
| WO2021007812A1 (en) * | 2019-07-17 | 2021-01-21 | 深圳大学 | Deep neural network hyperparameter optimization method, electronic device and storage medium |
| CN113128119A (en) * | 2021-04-21 | 2021-07-16 | 复旦大学 | Filter reverse design and optimization method based on deep learning |
| WO2023038070A1 (en) * | 2021-09-10 | 2023-03-16 | Mitsubishi Electric Corporation | Multiplier-less sparse deep neural network |
| US20230081531A1 (en) * | 2021-09-10 | 2023-03-16 | Mitsubishi Electric Research Laboratories, Inc. | Multiplier-Less Sparse Deep Neural Network |
Non-Patent Citations (1)
| Title |
|---|
| XING QUAN 等: "A Low-Loss and High-Isolation Transformer-Based mm-Wave SPDT with Integrated Fan-out Wafer Level Packaging", JOURNAL OF CIRCUITS, SYSTEMS AND COMPUTERS, vol. 29, no. 07, 24 September 2019 (2019-09-24) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN119720906B (en) | 2025-08-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113128119B (en) | Filter Reverse Design and Optimization Method Based on Deep Learning | |
| CN110414565B (en) | Group Lasso-based neural network cutting method for power amplifier | |
| CN111859795B (en) | Polynomial aided neural network behavior modeling system and method for power amplifier | |
| CN115033204B (en) | High-energy-efficiency approximate multiplier with reconfigurable precision and bit width | |
| CN107947761A (en) | Change threshold percentage renewal adaptive filter algorithm based on lowest mean square quadravalence | |
| CN109858616A (en) | Power amplifier behavior level modeling system neural network based and method | |
| CN116345469A (en) | Power grid power flow adjustment method based on graph neural network | |
| CN107274009A (en) | A kind of time series data multistep forecasting method and system based on correlation | |
| CN114841106A (en) | Integrated circuit optimization method and system based on rule-guided genetic algorithm | |
| CN116663476A (en) | A High-Order High-Temperature Superconducting Filter and Its Circuit Topology Rapid Design Method | |
| CN116992806A (en) | Automatic optimization method for large-scale analog integrated circuit based on self-attention mechanism | |
| CN119720906B (en) | A method, system, device and medium for optimizing analog equalizer design based on deep learning algorithm | |
| CN117992567A (en) | Method, system and storage medium for optimizing transducer model based on dynamic window | |
| CN119277491B (en) | Base station energy-saving regulation and control method based on diffusion model and reinforcement learning | |
| CN109902807A (en) | A Distributed Thermal Modeling Method for Many-Core Chips Based on Recurrent Neural Networks | |
| CN116488970B (en) | Balancing method for high-speed optical communication system | |
| CN117997379A (en) | Multi-input multi-output power line carrier noise modeling method based on differentiated selection | |
| CN118214629A (en) | A Sparse Underwater Acoustic Channel Estimation Method Based on Proportional Recursive P-Power Error Criterion | |
| CN114462232B (en) | A microwave device modeling method based on multi-layer neural network model structure adaptation | |
| CN113076699B (en) | Antenna optimization method based on Bayesian optimization of multi-output Gaussian process | |
| CN114048550B (en) | Vehicle cab sound quality evaluation method of PSO optimization delay operator model | |
| CN113516170A (en) | Image classification method based on Bayesian neural network random addition saturation structure | |
| CN119808946A (en) | A high-performance inference circuit implementation method for a general artificial intelligence nonlinear activation function | |
| CN112134631A (en) | Method and system for evaluating average capacity of wireless communication system | |
| TW202030647A (en) | System and method for reducing computational complexity of artificial neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |