A method, system, device and medium for optimizing analog equalizer design based on deep learning algorithm

Abstract

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 using the electromagnetic response curve of the equalizer circuit. Additional equalizer circuit parameter data is generated using the trained inverse neural network and combined with the original simulation data into a new data set to solve the problem of excessive simulation software time consumption. A forward neural network is trained using this new data set. The trained forward neural network outputs a predicted equalizer electromagnetic response curve, which is compared with a target equalizer electromagnetic response curve. If 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

[0001] The present invention belongs to the technical field of analog equalizers, and in particular relates to an analog equalizer optimization design method, system, device and medium based on a deep learning algorithm. Background Art

[0002] In the fields of wireless communications and precision measurement, system performance is often constrained by non-ideal factors such as the link's amplitude-frequency response and group delay characteristics. These non-ideal characteristics are particularly significant when the system operates within a wide frequency band, resulting in a serious negative impact on the integrity and accuracy of signal transmission. In the field of precision measurement, solid-state power amplifiers are core components, and their design standards play a decisive role in overall performance. However, due to inherent nonlinear characteristics of the devices or problems such as the multi-stage cascade of power circuits, the uneven transmission parameters of two-port and multi-port devices can easily induce signal distortion, resulting in large fluctuations in the system's output amplitude over a wide frequency band, leading to waveform degradation in the amplifier circuit, causing inter-symbol interference, and ultimately affecting the system's communication quality.

[0003] To overcome these effects, the industry's current strategy is to introduce equalizers at key nodes. The equalizer's primary function is to compensate for component and channel nonuniformity and unevenness, reducing signal distortion and improving output power stability. This optimization can significantly improve system performance in broadband applications, enhancing signal clarity and reliability.

[0004] However, with the advancement of communications technology, the demand for wideband and high-gain characteristics in power devices continues to grow. This requires considering more target parameters and design criteria during the design process, increasing the difficulty of equalizer design. Especially in the design of microwave power amplifiers, when the device structure becomes highly complex, the equalizer design and optimization process can become extremely time-consuming. Using traditional design methods, completing an equalizer design can take hours to days. In recent years, with the application of artificial intelligence in an increasing number of fields, neural networks have shown excellent robustness and generalization capabilities, making them suitable for rapid design under different indicators. Therefore, introducing neural network algorithms into the design process can facilitate multi-objective and multi-indicator design.

[0005] Existing implementation solutions:

[0006] 1. The traditional circuit design process involves selecting a suitable equalizer circuit structure based on specified target indicators, manually calculating the system function expression of the circuit based on the circuit structure, and calculating the relevant parameter values ​​of the circuit based on relevant requirements. This method is not only time-consuming and labor-intensive, but also difficult to ensure the quality of the design due to manual calculations.

[0007] 2. Neural network microwave design process. Currently, filters are usually designed using neural networks. This method is currently limited to filters in specific frequency bands. First, the filter structure needs to be specified. A large amount of training data and computing resources are required to achieve good results. The model may need to be retrained for different application scenarios.

[0008] 3. In the existing neural network training method, too few data sets will cause the neural network to fall into a local optimal solution due to the influence of non-unique solutions. However, if the data set is large, it will take too long to obtain the data set through simulation software. In addition, optimization algorithms such as particle swarm optimization have high requirements for initial values. Simulation software needs to be used continuously during the iteration process, which is also time-consuming.

[0009] The shortcomings of the existing technology mainly include:

[0010] 1. High cost and resource consumption: Traditional analog equalizer design methods rely on complex electromagnetic simulation and difficult computational design, which not only increases design costs but also requires a lot of computing resources.

[0011] 2. Inefficiency and long cycle: The manual design and adjustment process is time-consuming and inefficient, especially when faced with complex circuit designs. It requires a large number of manual iterations and experiments, which prolongs the design cycle.

[0012] 3. Design complexity: As communication systems demand high-performance channel characteristics, equalizer design becomes more complex, and traditional design methods are unable to cope with the ever-increasing design standards and target parameters.

[0013] 4. Dependence on professional knowledge: Existing design methods rely on the professional knowledge and experience of the designer, which makes it difficult for ordinary people to quickly master and apply them to actual designs.

[0014] Patent application publication number CN 113128119 A discloses a deep learning-based filter inverse design and optimization method. The filter's structural parameters are designed based on changes in its electromagnetic response. The inverse network is trained using data generated by a forward network. The filter's electromagnetic response curve is synthesized using Chebyshev polynomials, ultimately outputting the optimized filter response curve and filter structural parameters. However, because this method uses Chebyshev polynomials to synthesize the filter's electromagnetic response, it requires manual selection of Chebyshev polynomials of appropriate order. High-order Chebyshev polynomials can lead to significant errors, resulting in reduced accuracy and wasted computational costs. Summary of the Invention

[0015] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide an analog equalizer optimization design method, system, device and medium based on a deep learning algorithm. The analog equalizer circuit is designed and optimized through a deep learning method, and an inverse neural network is used to self-generate additional equalizer circuit parameter data. A new data set is formed with the original simulation data for training a forward neural network, and finally the equalizer circuit parameters predicted by the inverse neural network model and the equalizer electromagnetic response curve predicted by the forward neural network are obtained, thereby realizing the automatic design and optimization of the analog equalizer circuit and improving the design efficiency and accuracy.

[0016] In order to achieve the above object, the technical solution adopted by the present invention is:

[0017] A method for optimizing the design of an analog equalizer based on a deep learning algorithm comprises the following steps:

[0018] Step 1: Using a randomly generated electromagnetic response curve of the equalizer circuit as an input to the inverse neural network to train the inverse neural network, thereby obtaining a trained inverse neural network, and using the trained inverse neural network to generate additional equalizer circuit parameter data;

[0019] Step 2: Combine the additional equalizer circuit parameter data generated by the inverse neural network trained in step 1 with the original simulation data to form a new data set. The new data set is used to train the forward neural network. After reaching the convergence condition, a trained forward neural network is obtained.

[0020] Step 3. Input the expected electromagnetic response curve of the equalizer circuit into the reverse neural network trained in step 1, generate predicted equalizer circuit parameter data through the reverse neural network, and then input the generated predicted equalizer circuit parameter data into the forward neural network trained in step 2. The forward neural network outputs the predicted actual electromagnetic response curve of the equalizer circuit; finally, input the predicted actual electromagnetic response curve of the equalizer circuit into the reverse neural network trained in step 1 to obtain the actual equalizer circuit parameter data; finally, obtain the actual equalizer circuit parameter data predicted by the reverse neural network and the actual equalizer circuit electromagnetic response curve predicted by the forward neural network.

[0021] The inverse neural network includes an input layer, four fully connected hidden layers and corresponding Dropout layers and LeakyReLU activation functions, and an output layer, wherein the number of neurons in the four fully connected hidden layers is set to 251, 128, 64 and 10 respectively. The training goal of the inverse neural network is to minimize the loss function between the predicted equalizer circuit parameter data and the actual equalizer circuit parameters.

[0022] The forward neural network includes an input layer, a fully connected hidden layer containing a LeakyReLU activation function, a hidden layer containing a Dropout layer and a LeakyReLU activation function, and an output layer, wherein the number of neurons in the input layer is set to 11, and the number of neurons in the hidden layer is set to 80. The training goal of the forward neural network is to minimize the loss function of the return loss and insertion loss curve of the predicted electromagnetic response curve and the return loss and insertion loss curve of the electromagnetic response curve corresponding to the equalizer circuit parameters in the original data set.

[0023] The present invention also provides an analog equalizer optimization design system based on a deep learning algorithm, comprising:

[0024] an additional equalizer circuit parameter data generation module, configured to use a randomly generated electromagnetic response curve of the equalizer circuit as an input to an inverse neural network to train the inverse neural network, obtain a trained inverse neural network, and generate additional equalizer circuit parameter data using the trained inverse neural network;

[0025] A forward neural network training module is used to combine the additional equalizer circuit parameter data generated by the trained inverse neural network with the original simulation data to form a new data set. The new data set is used to train the forward neural network. After reaching the convergence condition, a trained forward neural network is obtained.

[0026] The actual equalizer circuit parameter data and electromagnetic response curve generation module is used to input the expected equalizer circuit electromagnetic response curve into the trained reverse neural network, generate predicted equalizer circuit parameter data through the reverse neural network, and then input the generated predicted equalizer circuit parameter data into the trained forward neural network, and the forward neural network outputs the predicted actual equalizer circuit electromagnetic response curve; finally, the predicted actual equalizer circuit electromagnetic response curve is input into the reverse neural network trained in step 1 to obtain the actual equalizer circuit parameter data; finally, the actual equalizer circuit parameter data predicted by the reverse neural network and the actual equalizer circuit electromagnetic response curve predicted by the forward neural network are obtained.

[0027] The present invention also provides an analog equalizer optimization design device based on a deep learning algorithm, comprising:

[0028] Memory: a computer-readable device storing a computer program for the above-mentioned analog equalizer optimization design method based on a deep learning algorithm;

[0029] Processor: used to implement the analog equalizer optimization design method based on deep learning algorithm when executing the computer program.

[0030] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement the analog equalizer optimization design method based on a deep learning algorithm.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. Reduce manual calculation costs.

[0033] The present invention adopts a design method that combines a forward neural network with an inverse neural network. Part of the additional circuit parameter data generated by the trained inverse neural network is combined with the original simulation data to form a new data set. The new data set is used to train the forward neural network. The trained inverse neural network can generate equalizer circuit parameters that meet the design requirements. The trained forward neural network can generate an equalizer electromagnetic response curve that meets the design requirements. In this way, arbitrarily complex equalizing circuit design can be realized, and the design of any equalizing effect can basically be achieved within the frequency range of 0-10GHz. Accurate design results can be obtained in a relatively short time, avoiding tedious manual calculations and multiple iterations.

[0034] 2. Save computing resources and design time

[0035] The present invention uses a forward neural network to predict the electromagnetic response curve of the equalizer and a reverse neural network to predict the equalizer circuit parameters. It only requires a small number of random equalizer circuit parameters and simulated electromagnetic response curves as input data. The reverse neural network can automatically generate a large number of new equalizer circuit electromagnetic simulation results, saving the time cost and computing resources of repeatedly calling electromagnetic simulation software.

[0036] 3. Improve design efficiency.

[0037] The present invention adopts a design method combining an inverse neural network and a forward neural network. The inverse neural network can automatically generate a large amount of data for training the forward neural network using a small amount of initial data set, which not only increases the size of the data set, but also avoids the large amount of time consumption caused by repeatedly calling the simulation software, greatly reducing the model training time and design cycle. In addition, the trained forward neural network has a strong response speed. For one design, it only takes 1.35 seconds, and for 1000 random different designs, it only takes 2.67 seconds. When it comes to equalizers with a large number of different design indicators, the equalizer design results can be output in a very short time, thereby greatly improving the design efficiency.

[0038] 4. Reduce design complexity

[0039] By automating the design of complex equalizer circuits through deep learning models, the present invention can meet the design requirements of complex equalizers with high performance, wide bandwidth, and multiple design indicators, so that circuit design can still be completed accurately and efficiently when facing increasingly complex communication system design standards.

[0040] 5. Simplify design difficulty

[0041] This invention simplifies traditional calculation and optimization processes by automatically generating design results through a deep learning algorithm. With a trained neural network, users only need to provide the corresponding design specifications to design a qualified equalizer circuit. This allows non-experts to quickly get started and achieve efficient equalizer circuit design, reducing reliance on specialized knowledge.

[0042] 6. Optimize design accuracy

[0043] The present invention utilizes the excellent fitting effect of neural networks for nonlinear functions to realize the design of arbitrarily complex equalization circuits, and the corresponding electromagnetic response curve return loss and insertion loss curves are well consistent with the actual results, with design errors less than 1dB.

[0044] 7. Be forward-looking

[0045] This paper utilizes deep learning to design passive equalizers. This allows for the exploration of a broader design parameter space, uncovering innovative design solutions that might be overlooked by traditional design methods. As a forward-looking technical solution, this paper can meet the increasingly complex and demanding requirements for various equalizers in future experimental scenarios.

[0046] In summary, the present invention designs the equalizer circuit through a deep learning method, adopts a method that combines an inverse neural network with a forward neural network, and automatically generates a large amount of data for training, saving labor costs and time consumption. While ensuring that the design accuracy meets the requirements, it has the advantages of reducing manual calculation costs, improving design efficiency and design accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the overall framework diagram of the equalizer design method of the present invention.

[0048] Figure 2 It is a structural diagram of the equalizer circuit of the present invention.

[0049] Figure 3 It is the reverse neural network structure diagram of the present invention.

[0050] Figure 4 It is a structural diagram of the forward neural network of the present invention.

[0051] Figure 5 This is a graph showing the return loss curve prediction results of the present invention.

[0052] Figure 6 This is a diagram showing the insertion loss curve prediction results of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0054] The present invention addresses the common problems existing in today's integrated design methods for analog passive equalizer circuits, especially when their application scenarios gradually present increasingly wider operating frequency bands, increasingly larger required compensation gains, and more and more design requirements such as application index requirements. Traditional integrated design methods for equalizer circuits rely on manual calculations and continuous adjustments. The method of calculating transfer functions and then selecting circuit parameters is not only time-consuming and complex in the calculation process, but also becomes particularly difficult as the design indicators become increasingly complex.

[0055] In order to solve this problem, the present invention proposes an innovative method that can easily realize the design of equalizers for various complex indicators through the use of a two-layer neural network. The method first trains the inverse neural network through the electromagnetic response curve of the equalizer circuit, generates additional equalizer circuit parameter data through the trained inverse neural network, and combines it with the original simulation data into a new data set to solve the problem of the simulation software taking too long. The forward neural network is trained with this new data set, and the trained forward neural network outputs a predicted equalizer electromagnetic response curve, which is compared with the target equalizer electromagnetic response curve. If the conditions are met, the inverse neural network is used to generate the actual equalizer circuit parameters that meet the conditions, and the design is completed. 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.

[0056] The present invention solves the following technical problems:

[0057] 1. Comprehensive circuit design method for complex indicators: To solve the high-cost and high-workload problems of traditional design methods that rely on manual calculations and continuous optimization and adjustment, the present invention needs to automatically design the required balanced circuit, including selecting the appropriate balanced circuit structure.

[0058] 2. High efficiency and high precision when designing a large number of complex circuits: When there are a large number of equalizers to be designed, the traditional method requires a lot of time and computing resources. The present invention can complete a large number of circuit designs in a very short time, achieving high efficiency and high precision design.

[0059] The following takes the design of a T-type equalizer circuit as an example for further detailed explanation.

[0060] like Figure 1As shown, an analog equalizer optimization design method based on a deep learning algorithm includes the following steps:

[0061] Step 1: Generate additional equalizer circuit parameter data using an inverse neural network;

[0062] 1,000 sets of randomly generated electromagnetic response curves of equalizer circuits are used as input to train the inverse neural network. The output of the inverse neural network is the equalizer circuit parameter data. The training goal is to minimize the loss function between the predicted equalizer circuit parameter data and the actual equalizer circuit parameters. After reaching the convergence condition or the maximum number of iterations, a trained inverse neural network is obtained, and then the trained inverse neural network is used to generate additional equalizer circuit parameter data. The inverse neural network can self-generate more equalizer circuit parameter data, making up for the large amount of time consumed by calling simulation software to generate data.

[0063] Figure 2 This is a schematic diagram of the equalizer circuit structure used in the present invention, wherein the lumped parameter circuit adopts a broadband network model combining broadband series and parallel connections.

[0064] The ten key circuit parameters of the equalizer are C1, C2, C3, L1, L2, L3, R1, R2, R3, and R4. A suitable parameter range is set, and within this range, 1,000 return loss and insertion loss curves are generated using electromagnetic simulation software. The frequency of the electromagnetic response curve is 3-10 GHz, and the return loss must be less than -10 dB at 3-10 GHz. Each curve has 251 data points within this frequency range. These 1,000 sets of samples are made into a data set, 900 of which are training sets and 100 are test sets. Figure 3 The trained inverse neural network automatically generates 3,600 sets of data, which together with the previous 900 sets of arrays form 4,500 sets of data.

[0065] like Figure 3 As shown in the figure, the inverse neural network includes an input layer, four fully connected hidden layers and corresponding Dropout layers and LeakyReLU activation functions, and an output layer. The number of neurons in the four fully connected hidden layers are 251, 128, 64, and 10, respectively. The Dropout value is 0.07. The loss function is the MSR mean square error function. The maximum number of iterations is 1000. The training goal is to minimize the loss function between the predicted equalizer circuit parameter data and the actual equalizer circuit parameters.

[0066] Step 2: Combine the additional equalizer circuit parameter data generated by the inverse neural network trained in step 1 with the original simulation data to form a new data set. The new data set is used to train the forward neural network. After reaching the convergence condition, a trained forward neural network is obtained. The forward neural network is used to generate a predicted electromagnetic response curve of the equalizer circuit.

[0067] Specifically, the 4,500 data sets generated in step 1 were used to train a forward neural network. This neural network, by providing predicted frequency information, can obtain the return loss and insertion loss at corresponding frequencies. After repeating this process 251 times, the return loss and insertion loss curves predicted by the forward neural network for the 3-10 GHz frequency range were obtained.

[0068] like Figure 4 As shown, the forward neural network includes an input layer, a fully connected hidden layer containing a LeakyReLU activation function, a hidden layer containing a Dropout layer and a LeakyReLU activation function, and an output layer. The input layer has 11 neurons, and the hidden layers each contain 80 neurons. The loss function is the MSR mean square error function, the maximum number of iterations is 1000, and the training goal is to minimize the loss function of the predicted electromagnetic response curve return loss and insertion loss curve and the electromagnetic response curve return loss and insertion loss curve of the corresponding equalizer circuit parameters in the original data set.

[0069] Step 3. After the forward neural network training is completed, the expected electromagnetic response curve of the equalizer circuit is input into the reverse neural network trained in step 1, and the predicted equalizer circuit parameter data is generated by the reverse neural network. The generated predicted equalizer circuit parameter data is then input into the forward neural network trained in step 2, and the forward neural network outputs the predicted actual electromagnetic response curve of the equalizer circuit; the predicted actual electromagnetic response curve of the equalizer circuit is compared with the expected electromagnetic response curve of the equalizer circuit. If the error is large, the forward neural network is optimized, and the optimization target is the difference between the predicted electromagnetic response curve and the target electromagnetic response curve. When the loss function reaches a threshold or reaches the maximum number of iterations, the optimization stops. Finally, the predicted actual electromagnetic response curve of the equalizer circuit is input into the reverse neural network trained in step 1 to obtain the actual equalizer circuit parameter data; finally, the actual equalizer circuit parameter data predicted by the reverse neural network and the actual electromagnetic response curve of the equalizer circuit predicted by the forward neural network are obtained.

[0070] The error between the predicted electromagnetic response curve of the equalizer circuit generated in step 2 and the required target electromagnetic response curve of the equalizer is used as the optimization target. The loss functions of the forward neural network and the reverse neural network are both MSE mean square error functions. When the loss function reaches the threshold or the maximum number of iterations is reached, the optimization stops. At this time, the output of the reverse neural network in step 1 is the circuit parameters of the designed equalizer, and the output of the forward neural network in step 2 is the electromagnetic response curve of the designed equalizer circuit.

[0071] Figure 5 、 Figure 6 This is the design result of the equalizer. Figure 5 The horizontal axis is frequency, and the vertical axis is dB(S11). The design goal is to keep dB(S11) < -10dB within the 5GHz-10GHz range. The target curve is shown by the solid line in the figure. The dashed line in the figure shows the results of the equalizer circuit parameters predicted by the inverse neural network in the simulation software. Figure 6 The horizontal axis is frequency, and the vertical axis is dB(S21). The goal is to achieve a monotonically increasing dB(S21) between 5 GHz and 10 GHz, reaching approximately 0 dB at 10 GHz. The target curve is shown by the solid line. The dashed line shows the simulation results of the equalizer circuit parameters predicted by the inverse neural network.

[0072] Figure 5 The horizontal axis is frequency, and the vertical axis is dB(S11). The goal is to keep dB(S11) < -10dB between 5GHz and 10GHz. The target curve is shown by the solid line. The dashed line shows the circuit parameter predictions from the inverse neural network as seen in the simulation software.

[0073] Figure 6 The horizontal axis is frequency, and the vertical axis is dB(S21). The goal is to achieve a monotonically increasing dB(S21) between 5 GHz and 10 GHz, reaching approximately 0 dB at 10 GHz. The target curve is shown by the solid line. The dashed line shows the circuit parameter predictions from the inverse neural network as seen in the simulation software.

[0074] Depend on Figure 5 、 Figure 6 It can be seen that the predicted results of the equalizer designed by the present invention match the target curves well. The mean absolute error of the return loss curve is 0.9665dB, and the mean absolute error of the insertion loss curve is 0.2447dB. The design errors are both less than 1dB, meeting the design requirements. This shows that after the neural network is trained, the user only needs to provide the corresponding relevant indicators. For example, to design an equalizer in the 5-10GHz range, the present invention only needs to provide the frequency range and the changes in the two electromagnetic response curves to design an equalizer circuit of any complexity.

[0075] The key points and protection points of the present invention are:

[0076] 1. Automated equalizer design based on deep learning model: The circuit parameter design of the equalizer is carried out. The equalizer indicators are reflected by the target balanced electromagnetic response curve. The present invention adopts forward and reverse two-layer neural networks for design to realize the automated design of analog equalizers, significantly improving the design efficiency and accuracy.

[0077] 2. The inverse neural network consists of an input layer, four fully connected hidden layers with corresponding Dropout layers and LeakyReLU activation functions, and an output layer. The four fully connected hidden layers have 251, 128, 64, and 10 neurons, respectively. The Dropout value is 0.07 for each layer. The loss function is the mean square error (MSR) function, with a maximum number of iterations of 1000. The training objective is to minimize the loss function between the predicted equalizer circuit parameter data and the actual equalizer circuit parameters. The success of the network training is determined by examining the convergence of the loss functions on the training and test sets, as well as the network's prediction performance on new datasets. If the trained network performs poorly, hyperparameters such as the dataset size, training period, and loss function need to be adjusted.

[0078] 3. The forward neural network consists of an input layer, a fully connected hidden layer with a LeakyReLU activation function, a hidden layer with a Dropout layer and a LeakyReLU activation function, and an output layer. The input layer has 11 neurons, and each hidden layer has 80 neurons. The loss function is the mean square error (MSR) function, and the maximum number of iterations is 1000. The training objective is to minimize the loss function between the predicted electromagnetic response curve return loss and insertion loss curves and the electromagnetic response curve return loss and insertion loss curves corresponding to the equalizer circuit parameters in the original dataset. The success of the network training is determined by examining the convergence of the loss function on the training and test sets, as well as the network's prediction performance on new datasets. If the trained network performance is poor, hyperparameters such as the dataset size, training period, and loss function need to be adjusted.

[0079] This paper proposes an optimized analog equalizer design method based on a deep learning algorithm, aiming to address the time-consuming, inefficient, and costly issues of traditional design methods in complex circuit design. By using a neural network model, the paper achieves automated and intelligent design of analog equalizer circuits. Through deep learning, the paper utilizes a two-layer neural network to achieve automated analog equalizer design.

[0080] The present invention also provides an analog equalizer optimization design system based on a deep learning algorithm, comprising:

[0081] an additional equalizer circuit parameter data generation module, configured to implement the step 1 of using the randomly generated electromagnetic response curve of the equalizer circuit as an input to the inverse neural network to train the inverse neural network, thereby obtaining a trained inverse neural network, and generating additional equalizer circuit parameter data using the trained inverse neural network;

[0082] A forward neural network training module is used to combine the additional equalizer circuit parameter data generated in step 2 using the inverse neural network trained in step 1 with the original simulation data to form a new data set. The new data set is used to train the forward neural network. After reaching the convergence condition, a trained forward neural network is obtained.

[0083] The actual equalizer circuit parameter data and electromagnetic response curve generation module is used to implement the steps of inputting the expected equalizer circuit electromagnetic response curve into the reverse neural network trained in step 1 in step 3, generating predicted equalizer circuit parameter data through the reverse neural network, and then inputting the generated predicted equalizer circuit parameter data into the forward neural network trained in step 2, and the forward neural network outputting the predicted actual equalizer circuit electromagnetic response curve; finally, inputting the predicted actual equalizer circuit electromagnetic response curve into the reverse neural network trained in step 1 to obtain the actual equalizer circuit parameter data; and finally obtaining the actual equalizer circuit parameter data predicted by the reverse neural network and the actual equalizer circuit electromagnetic response curve predicted by the forward neural network.

[0084] The present invention also provides an analog equalizer optimization design device based on a deep learning algorithm, comprising:

[0085] Memory: a computer-readable device storing a computer program for the above-mentioned analog equalizer optimization design method based on a deep learning algorithm;

[0086] Processor: used to implement the analog equalizer optimization design method based on deep learning algorithm when executing the computer program.

[0087] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement the analog equalizer optimization design method based on a deep learning algorithm.

Claims

1. A method for optimizing the design of an analog equalizer based on a deep learning algorithm, characterized in that: The following steps are involved: Step 1: Using a randomly generated electromagnetic response curve of the equalizer circuit as an input to the inverse neural network to train the inverse neural network, thereby obtaining a trained inverse neural network, and using the trained inverse neural network to generate additional equalizer circuit parameter data; Step 2: Combine the additional equalizer circuit parameter data generated by the inverse neural network trained in step 1 with the original simulation data to form a new data set. The new data set is used to train the forward neural network. After reaching the convergence condition, a trained forward neural network is obtained. Step 3. Input the expected electromagnetic response curve of the equalizer circuit into the reverse neural network trained in step 1, generate predicted equalizer circuit parameter data through the reverse neural network, and then input the generated predicted equalizer circuit parameter data into the forward neural network trained in step 2. The forward neural network outputs the predicted actual electromagnetic response curve of the equalizer circuit; finally, input the predicted actual electromagnetic response curve of the equalizer circuit into the reverse neural network trained in step 1 to obtain the actual equalizer circuit parameter data; finally, obtain the actual equalizer circuit parameter data predicted by the reverse neural network and the actual equalizer circuit electromagnetic response curve predicted by the forward neural network.

2. The analog equalizer optimization design method based on deep learning algorithm according to claim 1, characterized in that: The inverse neural network includes an input layer, four fully connected hidden layers and corresponding Dropout layers and LeakyReLU activation functions, and an output layer, wherein the number of neurons in the four fully connected hidden layers is set to 251, 128, 64 and 10 respectively. The training goal of the inverse neural network is to minimize the loss function between the predicted equalizer circuit parameter data and the actual equalizer circuit parameters.

3. The analog equalizer optimization design method based on deep learning algorithm according to claim 1, characterized in that: The forward neural network includes an input layer, a fully connected hidden layer containing a LeakyReLU activation function, a hidden layer containing a Dropout layer and a LeakyReLU activation function, and an output layer, wherein the number of neurons in the input layer is set to 11, and the number of neurons in the hidden layer is set to 80. The training goal of the forward neural network is to minimize the loss function of the return loss and insertion loss curve of the predicted electromagnetic response curve and the return loss and insertion loss curve of the electromagnetic response curve corresponding to the equalizer circuit parameters in the original data set.

4. An analog equalizer optimization design system based on a deep learning algorithm based on the method according to any one of claims 1 to 3, characterized in that: include: an additional equalizer circuit parameter data generation module, configured to use a randomly generated electromagnetic response curve of the equalizer circuit as an input to an inverse neural network to train the inverse neural network, obtain a trained inverse neural network, and generate additional equalizer circuit parameter data using the trained inverse neural network; A forward neural network training module is used to combine the additional equalizer circuit parameter data generated by the trained inverse neural network with the original simulation data to form a new data set. The new data set is used to train the forward neural network. After reaching the convergence condition, a trained forward neural network is obtained. The actual equalizer circuit parameter data and electromagnetic response curve generation module is used to input the expected equalizer circuit electromagnetic response curve into the trained reverse neural network, generate predicted equalizer circuit parameter data through the reverse neural network, and then input the generated predicted equalizer circuit parameter data into the trained forward neural network, and the forward neural network outputs the predicted actual equalizer circuit electromagnetic response curve; finally, the predicted actual equalizer circuit electromagnetic response curve is input into the reverse neural network trained in step 1 to obtain the actual equalizer circuit parameter data; finally, the actual equalizer circuit parameter data predicted by the reverse neural network and the actual equalizer circuit electromagnetic response curve predicted by the forward neural network are obtained.

5. An analog equalizer optimization design device based on deep learning algorithm, characterized in that: include: Memory: a computer-readable device storing a computer program for the analog equalizer optimization design method based on a deep learning algorithm according to any one of claims 1 to 3; Processor: used to implement the analog equalizer optimization design method based on deep learning algorithm as described in any one of claims 1-3 when executing the computer program.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, can implement the analog equalizer optimization design method based on a deep learning algorithm as described in any one of claims 1 to 3.

Images