ADC calibration method based on genetic algorithm

Abstract

The invention belongs to the field of ADC digital signal calibration, and particularly relates to an ADC calibration method based on a genetic algorithm, and the method comprises the steps: obtaining an input signal, inputting the input signal into an ADC model, and outputting a to-be-calibrated signal; inputting a signal to be calibrated into the trained calibration network to obtain a calibrated signal; the training of the calibration network comprises the following steps: constructing a non-ideal truth value data set; inputting data in the non-ideal truth value data set into the ADC model to obtain a non-ideal truth value signal to be calibrated; constructing an ADC error-free ideal truth value optimization problem according to the to-be-calibrated non-ideal truth value signal; solving an ADC error-free ideal truth value optimization problem by adopting a genetic algorithm to obtain an ADC ideal truth value; the ADC ideal true value is input into the calibration network for training, and when a convergence condition is met, training of the calibration network is completed; the method can be widely applied to various ADC architectures.

Description

Technical Field

[0001] The invention belongs to the field of ADC digital calibration algorithms, and in particular relates to an ADC calibration method based on a genetic algorithm. Background Art

[0002] An ADC (Analog-to-Digital Converter) is an electronic system or circuit that converts analog signals into digital signals. In modern electronic technology and communication systems, ADC plays a vital role because it allows computer systems, microcontrollers and other digital circuits to process information from analog sources such as sensors, audio equipment, and video signal sources. The operation of an ADC is divided into four main steps: first, sampling, which converts a continuous analog signal into a discontinuous analog signal; second, holding, which ensures that the sampled data is retained before the next sampling; these two steps are generally completed by a specific sample-and-hold circuit. The next steps, quantization and encoding, are responsible for converting the analog signal into a digital signal.

[0003] Due to the influence of factors such as device characteristics, manufacturing errors and environmental conditions, ADCs may have errors during the conversion process, and different types of ADCs or ADCs with different designs also produce different errors. Traditional calibration methods are usually only applicable to specific types of ADCs, and the types of errors that can be corrected are limited. The algorithm implementation is complex, the hardware cost is high, and it is not easy to transplant. In addition, these methods may also add additional circuit design to the analog part, which not only increases the design difficulty of the analog part, but also increases the uncertainty of performance.

[0004] Neural network-based calibration methods have been proposed in recent years and have achieved good results, but their training process relies on the true value of the signal, and the acquisition of the true value relies on an accurate ADC reference model or requires a priori information such as the frequency of a given signal for fitting. These methods will bring additional hardware overhead or introduce errors in actual use, and cannot be trained and updated online. Summary of the invention

[0005] In order to solve the problems existing in the above prior art, the present invention proposes an ADC calibration method based on a genetic algorithm, which includes: obtaining an input signal, inputting the input signal into an ADC model, and outputting a signal to be calibrated; inputting the signal to be calibrated into a trained calibration network to obtain a calibrated signal; training the calibration network includes: constructing a data set without ideal truth value; inputting data in the data set without ideal truth value into the ADC model to obtain a signal without ideal truth value to be calibrated; constructing an ADC error-free ideal truth value optimization problem according to the signal without ideal truth value to be calibrated; solving the ADC error-free ideal truth value optimization problem using a genetic algorithm to obtain an ADC ideal truth value; inputting the ADC ideal truth value into the calibration network for training, and completing the training of the calibration network when a convergence condition is met.

[0006] Beneficial effects of the present invention:

[0007] The present invention is widely used in various ADC architectures. Users do not need to have an in-depth understanding of the specific causes of errors or nonlinear characteristics, nor do they need to develop special calibration schemes for specific error types. The calibration method of the present invention is highly versatile and easy to use, and is easy to deploy and apply quickly in different environments. The present invention does not rely on additional hardware such as reference circuits, but obtains the true value of the signal through a genetic algorithm, which can achieve low-cost, high-precision, and fast training and calibration. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 Schematic diagram of the overall architecture of the ADC calibration method based on genetic evolution algorithm of the present invention;

[0009] Figure 2 It is a genetic algorithm iteration effect diagram of the present invention;

[0010] Figure 3 A schematic diagram of a customized neural network architecture for calibration of the present invention;

[0011] Figure 4 Spectrum diagrams before and after calibration of the present invention. DETAILED DESCRIPTION

[0012] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0013] A genetic algorithm-based ADC calibration method, the method comprising: obtaining an input signal, inputting the input signal into an ADC model, and outputting a signal to be calibrated; inputting the signal to be calibrated into a trained calibration network to obtain a calibrated signal; training the calibration network comprises: constructing a data set without ideal truth value; inputting data in the data set without ideal truth value into the ADC model to obtain a signal without ideal truth value to be calibrated; constructing an ADC error-free ideal truth value optimization problem according to the signal without ideal truth value to be calibrated; solving the ADC error-free ideal truth value optimization problem using a genetic algorithm to obtain an ADC ideal truth value; inputting the ADC ideal truth value into the calibration network for training, and completing the training of the calibration network when a convergence condition is met.

[0014] The overall architecture of the ADC calibration method based on the genetic evolution algorithm of this embodiment is as follows: Figure 1 As shown. During training, ADC outputs of various structures with various errors are first obtained, and then the corresponding ideal true values ​​are obtained using a genetic evolutionary algorithm and a complete data set is constructed to train the custom calibration network designed by the present invention. During testing, error signals are collected online and then input into the calibration network for calibration. The specific implementation plan is as follows:

[0015] S1. Construct a data set with multiple errors and multiple ADC structures without ideal true values: Establish a simulated ADC model and add errors to it, input time series and signals, obtain ADC simulated error output and collect it into a data set. Collect the measured error output data of multiple ADC architectures and merge them into the above data set.

[0016] In this embodiment, the errors include capacitor mismatch error, operational amplifier limited bandwidth error, memory effect, aperture jitter error, inter-channel gain mismatch error, and time mismatch error. The ADC structures include pipelined, SAR, Sigma-Delta, and Time-Interval. The ADC sampling frequency is 1G and the resolution is 12bit. For the input signal, we use single / dual / multi-tone signals with a frequency in the range of [40MHz, 450MHZ].

[0017] S2. Establish the framework of the ADC error-free ideal true value acquisition problem: transform the ADC ideal true value acquisition problem into a target optimization problem. The expression of the target optimization problem is:

[0018] y = argmin(L(x+dx))

[0019] Among them, x is the ADC error output, dx is the compensation amount to be optimized, and L is the objective function. This technical solution uses a widely used ADC indicator: the opposite number of the spurious free dynamic range SFDR as the objective function. The meaning of this optimization problem is to find the ideal compensation amount and then get the ideal true value, so that the ADC spurious free dynamic range is maximized.

[0020] S3. Use the genetic evolution algorithm to obtain the ideal true value of ADC, that is, to find the solution to the above optimization problem: given the initial solution of the problem, construct a Gaussian distribution and sample it multiple times, combine the sampled value with the initial solution and calculate its objective function value, calculate the variance and step size of the next sampling distribution according to the objective function value, and repeat the above steps until the convergence condition is met. The obtained true value is merged into the above data set in a one-to-one correspondence with the error data.

[0021] Specifically, based on the initial solution of the given optimization problem, a Gaussian distribution is generated, and samples are extracted from the distribution multiple times. The extracted sample values ​​are combined with the initial solution, and the obtained objective function value is calculated. Based on these objective function values, the variance and step size of the subsequent sampling distribution are adjusted. Then, the above process is repeated until the established convergence criterion is met. The iterative effect of the above genetic evolution algorithm in a specific embodiment is as follows: Figure 2 As shown in the figure, it can be seen that the inverse number of the objective function SFDR decreases continuously during the iteration process and reaches the ideal optimal point. At this time, the independent variable is the ideal true value corresponding to the desired ADC output value with error.

[0022] In this embodiment, the specific process of solving the ADC error-free ideal true value optimization problem using the genetic algorithm includes:

[0023] Step 1: Initial genetic algorithm parameters, including λ, ω i 、c c 、c σ ,d σ ,c1,c μ , μ, μ ω and σ0; set the number of iterations T; where λ is the number of samples, ω i 、c σ 、c σ ,d σ ,c1,c μ , μ, μ ω They are all iterative control parameters;

[0024] Step 2: Assign initial values ​​to the parameters, i.e., let Σ=I,P c =0,P σ =0, Δx=0 and σ=σ0; where I is the identity matrix, Σ is the covariance matrix, P c is the anisotropic evolution path, Pσ is the isotropic evolution path, Δx is the independent variable to be optimized, σ is the evolution step, and σ0 is the initial evolution step;

[0025] Step 3: Generate λ sample vectors e according to the covariance matrix i ;

[0026] Step 4: According to the sample vector e i Compute candidate solutions; their expressions are:

[0027] z i =Δx+σe i

[0028] Step 5: Evaluate each candidate solution z i The cost value L g (x+z i ), and the current iteration number increases by 1;

[0029] Step 6: Determine whether the current number of iterations reaches the maximum number of iterations. If yes, go to step 7; otherwise, return to step 3.

[0030] Step 7. Sort z and e according to cost value, and then let z i:λ and e i:λ Respectively represent the i-th vector after sorting;

[0031] Step 8: Calculate based on the sorted i-th vector:

[0032] Step 9: Update Δx, where the update formula is: Δx←Δx+σe i ;

[0033] Step 10: P c Update, where the update formula is:

[0034]

[0035] Step 11: P σ Update, where the update formula is:

[0036]

[0037] Step 12: Update the covariance matrix, where the update expression is:

[0038]

[0039] Step 13: Update σ, and its expression is:

[0040]

[0041] Where N is the independent variable dimension.

[0042] S4. Design of neural network structure for calibration: The present invention uses a custom designed neural network structure specifically for calibration tasks.

[0043] Specifically, Figure 3 The neural network architecture is designed as shown in Figure 1. The design of the architecture begins with a one-dimensional convolutional layer that serves as the main entry point for preprocessing the input data. This layer is crucial for feature extraction and prepares the data for subsequent processing. The data output from this initial convolutional layer is strategically split into three specialized branches: linear branch, nonlinear branch, and residual branch. Each branch has its own unique purpose and contributes to the overall performance of the network. The linear branch and the nonlinear branch are each equipped with two one-dimensional convolutional layers. The main difference between the two branches lies in the nature of their activation functions. The linear branch adopts a linear activation function to preserve the original characteristics of the data, while the nonlinear branch contains a nonlinear activation function to capture the complex relationships within the data. The three branches do not operate in isolation; instead, they are integrated into the network through a weighted sum operation. This integration is controlled by three learnable parameters that are adaptively updated during the training phase. Fine-tuning of these parameters provides an additional layer of optimization, thereby enhancing the network's ability to calibrate noisy data.

[0044] S5. The ADC error output obtained above and its corresponding ideal true value are used to train the network designed above, and finally an ADC calibration network with multiple error calibration capabilities is obtained for use in ADC calibration tasks.

[0045] In this embodiment, the true value obtained by the genetic algorithm is used to train the neural network, the number of training rounds is 300, the loss function is the mean square error MSE, and the learning rate is 0.01. Before starting the training, the network parameters are randomly initialized with a normal distribution.

[0046] The expression of mean square error is:

[0047] MSE(x,y)=∑(xy) 2

[0048] Where x is the signal value output by the ADC with error, and y is the target signal value.

[0049] Preferably, during the model training process, the number of iterations is set to 10,000 and the threshold is set to 0.01. When the model reaches the number of iterations or the target variable value after the iteration is less than the set threshold, the model converges.

[0050] The trained calibration network is used to calibrate the unknown input signal, and the spectrum result is as follows: Figure 4As shown, it can be seen that both spurious and harmonic waves are filtered out, and the maximum spurious signal drops from -76.82dB to -96.87dB.

[0051] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An ADC calibration method based on genetic algorithm, characterized in that: include: Obtain the input signal, input the input signal into the ADC model, and output the signal to be calibrated; Inputting the signal to be calibrated into the trained calibration network to obtain a calibrated signal; Training the calibration network includes: constructing a data set without ideal truth value; inputting the data in the data set without ideal truth value into the ADC model to obtain a signal without ideal truth value to be calibrated; constructing an ADC error-free ideal truth value optimization problem based on the signal without ideal truth value to be calibrated; solving the ADC error-free ideal truth value optimization problem using a genetic algorithm to obtain an ADC ideal truth value; inputting the ADC ideal truth value into the calibration network for training, and completing the training of the calibration network when the convergence condition is met.

2. The ADC calibration method based on genetic algorithm according to claim 1, characterized in that: The input signal is single / dual / multi-tone signal with a frequency in the range of [40MHz, 450MHZ].

3. The ADC calibration method based on genetic algorithm according to claim 1, characterized in that: The ADC model has multiple errors, including capacitor mismatch error, operational amplifier limited bandwidth error, memory effect, aperture jitter error, inter-channel gain mismatch error, and time mismatch error.

4. The ADC calibration method based on genetic algorithm according to claim 1, characterized in that: Constructing a data set without ideal truth value includes: obtaining time series and corresponding signals; inputting the time series and signals into ADC models with different framework structures to obtain error-bearing data of different ADC frameworks; and combining all data to obtain a data set without ideal truth value.

5. The ADC calibration method based on genetic algorithm according to claim 1, characterized in that: The ADC error-free ideal truth value optimization problem is: y = argmin(L(x+dx)) Where x is the ADC output with error, dx is the compensation amount to be optimized, and L is the objective function.

6. The ADC calibration method based on genetic algorithm according to claim 1, characterized in that: The genetic algorithm is used to solve the ADC error-free ideal true value optimization problem, including: given an initial solution to the problem, constructing a Gaussian distribution, and sampling the Gaussian distribution multiple times; combining the sampled values ​​with the initial solution and calculating its objective function value, calculating the variance and step size of the next sampling distribution according to the objective function value, and repeating the above steps until the convergence conditions are met.

7. The ADC calibration method based on genetic algorithm according to claim 1, characterized in that: The calibration network is a neural network model, which includes a linear branch, a nonlinear branch and a residual branch; The neural network model processes input data by: extracting features from the input data through a one-dimensional convolutional layer; inputting the extracted feature distribution into the linear branch, nonlinear branch, and residual branch for deep feature extraction; fusing the features extracted from the three branches to obtain fused features; and calibrating the input signal based on the fused features.

8. The ADC calibration method based on genetic algorithm according to claim 1, characterized in that: The convergence condition of the model is: the number of iterations and threshold are set during the model training process. When the model reaches the number of iterations or the target variable value after the iteration is less than the set threshold, the model converges.

Images