An adc calibration method based on a fully connected neural network

Abstract

The present application belongs to the field of ADC calibration algorithm, and particularly relates to an ADC calibration method based on a full connection neural network. The method mainly comprises the following steps: constructing a full connection neural network; injecting a harmonic wave to an ideal signal, prolonging a sampling time to generate a training data set through random sampling; performing ideal quantization on the training data set with a higher precision than that of a target ADC, taking the ideal quantization as an input to train the full connection neural network; and performing calibration processing on an actual ADC chip according to the trained full connection neural network. The present application has the advantages that the full connection neural network is used to calibrate the target ADC, the source of errors and the influence of nonlinearity do not need to be concerned, the method is suitable for various ADC structures, has the characteristics of strong universality, easy transplantation, reduced difficulty in ADC circuit design, and improved performance of the ADC chip.

Description

Technical Field

[0001] This invention belongs to the field of ADC calibration algorithms, specifically relating to an ADC calibration method based on a fully connected neural network. Background Technology

[0002] The basic workflow of an ADC consists of four stages: sampling, holding, quantization, and encoding. The sampling stage converts a continuously changing analog signal into a discretely changing analog signal. The holding stage stores the sampling results until the next sampling. These two stages are usually implemented by a sample-and-hold circuit. The quantization and encoding stages convert the analog signal into a digital signal and then into digital codewords through certain encoding rules, ultimately achieving the output of a digital signal.

[0003] Due to device characteristics, manufacturing process errors, environmental factors, etc., ADCs typically generate errors during operation, and the errors generated by different modules or architectures of the same type of ADC are different. Traditional calibration techniques are mostly applicable to a single type of ADC, targeting a limited number of error types. Furthermore, their algorithms are complex to implement, have high hardware overhead, poor portability, and require additional circuit design in the analog section, increasing the design difficulty and performance uncertainty of the analog section. Summary of the Invention

[0004] To address the shortcomings of traditional calibration algorithms, such as limited applicability to specific ADC types, narrow error types, and complex implementation, this invention proposes a neural network ADC calibration method for calibrating ADC errors. This method is based on signal frequency domain modeling and is applicable to various ADC architectures, including Pipelined, SAR, and Sigma-Delta.

[0005] The technical solution of this invention is as follows:

[0006] An ADC calibration method based on a fully connected neural network is proposed to directly calibrate the digital codewords output by the ADC, thereby reducing errors in the codewords. The main steps include:

[0007] S1. Constructing the training dataset: using an arithmetic sequence t m，n =t m，1 -(n-1) / f s Generate time series T m , where f s t is the sampling frequency of the ADC. m，n This represents the sampling time of the ADC, m represents the number of data samples to be generated, and n = 1, 2, 3...; T represents the sampling time of the ADC. m When the set size is met, a time point is randomly selected and defined as t. m，1 As the current sampling time, t m，1Using this as a baseline, a time series t is obtained by continuously searching forward for 49 sampling periods of the ADC. m，2 , t m，3 , ..., t m，50 At each sampling time, the input signal is sampled to obtain a basic output codeword, and the data samples are composed of the basic output codeword. All data samples constitute the training data matrix.

[0008] Specifically, in the arithmetic sequence t m，1 Let t be the current sampling time of the ADC, and t m，1 ~∪(0, 1000×2πf) max This satisfies the uniformity of the training data generation in phase, thereby achieving decoupling of frequency and phase.

[0009] Specifically, the basic output codeword S in the training data m，n From the main lobe of the signal Second harmonic of signal Third harmonic of the signal Up to the kth harmonic of the signal Composition, that is δ represents white noise.

[0010] The main lobe of the signal For the basic output codeword S m，n The value under ideal conditions is given by the formula. Sure, Indicates the amplitude of the main lobe of the signal and f m The frequency of the main lobe of the signal is given by f, and its normalized frequency satisfies f. m / f s ~∪(f min f max ).

[0011] The second harmonic of the signal This is the small signal generated by the main lobe of the signal at twice the frequency. This signal is the nonlinear error generated by the ADC during the sampling process, as given by the formula... Sure, Represents the amplitude of the second harmonic and This indicates the phase relationship between the second harmonic and the main lobe, and

[0012] The third harmonic of the signal This is a small signal generated by the main lobe of the signal at three times the frequency. This signal represents the nonlinear error generated by the ADC during the sampling process, as shown by the formula... Sure.

[0013] Similarly, the kth harmonic of the signal This is a small signal generated by the main lobe of the signal at k times the frequency. This signal is the nonlinear error generated by the ADC during the sampling process, as given by the formula... Sure.

[0014] Wherein, the normalized frequency f min and f max Signal amplitude and Harmonic order k, second harmonic amplitude and Third harmonic amplitude and And all the way up to the kth harmonic amplitude and All settings are configured as needed.

[0015] S2. Based on the accuracy of the target ADC, select an ideal ADC to perform ideal quantization of the training data.

[0016] Specifically, the ideal quantization process is defined by the formula Q = round(S) m，n The LSB ( / LSB) × LSB is determined, where LSB represents the minimum resolution of the ideal ADC; and the accuracy relationship between the target ADC and the ideal ADC is N. train =N real +2, where N real For the accuracy of the target ADC, N train This represents the accuracy of an ideal ADC.

[0017] S3. The ideally quantized data is input into the neural network for training. The neural network module is a two-layer fully connected neural network with 50 input nodes, 1024 intermediate nodes, and 1 output node; wherein, the 50 input nodes correspond to data samples. The current sampled value of the ADC and the previous 49 sampled values ​​are combined to form a time series signal as the network input; 1024 intermediate nodes are the perception layer of the neural network, which are used to learn the amplitude, frequency and phase information of the input time series, so as to analyze network errors and make corresponding compensations; one output node outputs the calibration value of the current sampled value of the ADC; after training with training data, the trained neural network is obtained.

[0018] Specifically, when initializing the neural network, the weights of the first layer need to be set to have a mean of 0 and a variance of 0. The values ​​follow a normal distribution with all biases equal to 0. The weights of the second layer satisfy the following conditions: mean 0 and variance 0. The model is a normal distribution with a bias of 0. The activation function is ReLU, the loss function is MSE, the optimization method is Adam, and the initial learning rate is 0.1.

[0019] S4. The output of the target ADC chip is calibrated using the trained neural network. Specifically, the signal sampling data of the target ADC chip is converted from serial to parallel and input into the trained neural network. The output data of the neural network is the calibrated signal.

[0020] The beneficial effects of this invention are as follows: This invention is applicable to various ADC structures, without needing to focus on the source of error and nonlinear effects, and without needing to derive a single calibration algorithm for specific errors. It has the advantages of strong universality and easy portability. This invention is a neural network calibration method, which is highly decoupled from the ADC. Therefore, the module of this invention can be integrated into the back-end signal processing chip, reducing the circuit complexity and design difficulty of the ADC chip. Attached Figure Description

[0021] The accompanying drawings below help to better understand the following description of different embodiments of the invention, which schematically illustrate the main features of some implementations of the invention. These drawings and embodiments provide some embodiments of the invention in a non-limiting and non-exhaustive manner. For simplicity, the same or similar components or structures having the same function in different drawings are given the same reference numerals.

[0022] Figure 1 This is a schematic diagram of the overall architecture of a neural network ADC calibration method proposed in this invention.

[0023] Figure 2 A flowchart illustrating the workflow for generating the training dataset proposed in this invention.

[0024] Figure 3 This invention addresses the phase inhomogeneity problem arising from signal frequency variations during training dataset generation.

[0025] Figure 4 This is a flowchart illustrating a specific implementation scheme for training a neural network proposed in this invention.

[0026] Figure 5 The calibration effect of the neural network trained for a specific example of the present invention on actual Pipelined and Sigma-Delta structure ADCs. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0028] The overall architecture of the neural network ADC calibration method in this embodiment is as follows: Figure 1 As shown, the ideal signal in the training dataset is injected into an ideal ADC. The ADC can be of various structures such as Pipelined, SAR, and Sigma-Delta. Simultaneously, the high-order harmonic error of the signal is injected into the ADC to generate corresponding sampled output data. The neural network inputs the received data vector into a fully connected network, calculates the difference between the network output and the ideal signal using the ReLU function, and feeds it back to the network for iterative optimization using a backpropagation mechanism. The specific implementation scheme is as follows:

[0029] B1, such as Figure 2 The training data is generated as shown, and the normalization frequency f is set. min =0.28 and f max =0.32, signal amplitude and ADC bit width N real =12, Total data volume m max =5000000, harmonic order k=3, second harmonic amplitude and Third harmonic amplitude and After setting the parameters, randomize them within a pre-defined range in each iteration and generate the ADC sampling time series T. m That is, in [0, 1000×2πf max Randomly select a time point t within the time range. m，1 Using the current sampling time as a reference, the time series t is obtained by continuously searching forward for 49 ADC sampling periods. m，2 , t m，3 , ..., t m，50 .

[0030] The present invention proposes that using random initial sampling points over a sufficiently long time range can solve the problem of phase non-uniformity in the sampled signal caused by continuous frequency variation, such as... Figure 3As shown, if sampling is performed within the period corresponding to the signal frequency in the first cycle, and sampling is still performed within the signal period of the first cycle when the signal frequency changes in the second cycle, some phases will be either not sampled or oversampled. When the sampling time is extended sufficiently, this non-uniform phase will become negligible, and the problem of phase non-uniformity will be solved.

[0031] Then based on the time series T m The basic output codeword S is obtained by performing a cyclic operation on 50 time points. m，n Each cycle obtains the value of the main lobe of the signal or the values ​​of each harmonic at the current moment.

[0032] Let's take the nth iteration as an example:

[0033] B1.1, t m，n Substitute main lobe function The desired output codeword is obtained.

[0034] B1.2, t m，n Substituting the second harmonic function The second-order nonlinear error is obtained, and then the third harmonic function is substituted. Three nonlinear errors were obtained.

[0035] B1.3, Utilization The ideal signal and nonlinear error are combined to obtain the basic output codeword, where δ is a very small white noise used to improve the robustness of the neural network.

[0036] The 50 basic output codewords are combined into a vector to obtain one output codeword sequence. If a sufficient sequence of output codewords is obtained, the loop stops; otherwise, the loop continues until the required amount of data is reached.

[0037] B2. Quantize the training dataset using an ideal ADC, selecting a quantization bit width N. train =14.

[0038] B3. Inject the training dataset into the fully connected neural network for training, such as... Figure 4 The diagram shown is a flowchart of the neural network training process in this embodiment. The specific training scheme is as follows:

[0039] B3.1 The network structure is set as a two-layer fully connected network with 50 input nodes, 1024 intermediate nodes, and 1 output node. The weight matrix of the first layer satisfies the following conditions: mean 0 and variance . The values ​​follow a normal distribution with all biases equal to 0. The second-layer weight matrix satisfies a mean of 0 and a variance of 0. The elements are normally distributed with biases of 0. The activation function is ReLU, the loss function is MSE, the optimization method is Adam, and the initial learning rate is 0.1.

[0040] B3.2. Train the initialized neural network. Set the number of iterations (epochs) to 80 and the batch size to 10. Due to the random sampling during the generation of the training dataset, the data generated in each iteration is different. Therefore, in this example, it is equivalent to training the fully connected neural network with 400 million sets of data in only one iteration. In this example, the loss function value of the network is recorded after each iteration. The convergence of the loss function after all iterations indicates that the neural network has been successfully trained.

[0041] B4. The performance of the trained neural network is tested using actual test data from a real ADC chip. The method for evaluating the performance of the ADC calibration method involves performing FFT analysis on the signals before and after calibration, analyzing parameters such as the signal-to-noise ratio (SNDR), spurious-free dynamic range (SFDR), and effective bits (ENOB) of the calibration signal on the spectrum. Specifically, unlike the random sampling of training data, this example selects 1024 sampling points for FFT analysis. The test data requires 1073 consecutive data samples. After calibration by the neural network, these data samples will generate 1024 consecutive digital signals, which are the calibration signals.

[0042] Furthermore, Figure 5 This paper demonstrates the calibration performance of the neural network trained in this example on actual Pipelined ADC and Sigma-Delta ADC chips. As shown in the figure, after processing by the neural network calibration method proposed in this invention, the SNDR of both ADC chips is improved by 8–9 dB, the SFDR by 12–15 dB, and the ENOB by 1–2 bits. Spectral clutter is effectively suppressed, indicating that the neural network ADC calibration method constructed in this embodiment has an effective calibration effect on actual ADCs.

[0043] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An ADC calibration method based on a fully connected neural network, characterized in that, Includes the following steps: S1. Obtain training data: using an arithmetic sequence t m，n =t m，1 -(n-1) / f s Generate time series T m , where f s t is the sampling frequency of the ADC. m，n This represents the sampling time of the ADC, m represents the number of data samples to be generated, and n = 1, 2, 3...; T represents the sampling time of the ADC. m When the set size is met, a time point is randomly selected and defined as t. m，1 As the current sampling time, t m，1 Using this as a baseline, a time series t is obtained by continuously searching forward for 49 sampling periods of the ADC. m，2 , t m，3 , ..., t m，50 At each sampling time, a basic output codeword is obtained, and the data sample is composed of the basic output codeword. All data samples constitute the training data. S2. Based on the accuracy of the target ADC, select an ideal ADC to perform ideal quantization on the training data; S3. Input the ideally quantized data into the neural network for training. The neural network module is a two-layer fully connected neural network with 50 input nodes, 1024 intermediate nodes, and 1 output node; where the 50 input nodes correspond to data samples. The current sampled value of the ADC and the previous 49 sampled values ​​are combined to form a time series signal as the network input; 1024 intermediate nodes are the perception layer of the neural network, which are used to learn the amplitude, frequency and phase information of the input time series, thereby analyzing network errors and making corresponding compensations; one output node outputs the calibration value of the current sampled value of the ADC. A trained neural network is obtained after training with training data; S4. Use the trained neural network to calibrate the output of the target ADC chip. Specifically, the signal sampling data of the target ADC chip output is converted from serial to parallel and input into the trained neural network. The output data of the neural network is the calibrated signal.

2. The ADC calibration method based on a fully connected neural network according to claim 1, characterized in that, Basic output codeword S m，n From the main lobe of the signal Second harmonic of signal Third harmonic of the signal Up to the kth harmonic of the signal Composition, that is δ represents white noise; The main lobe of the signal For the basic output codeword S m The value of n in the ideal case is given by the formula Sure, Indicates the amplitude of the main lobe of the signal and f m The frequency of the main lobe of the signal is given by f, and its normalized frequency satisfies f. m / f s ~∪(f min f max ); The second harmonic of the signal This is the small signal generated by the main lobe of the signal at twice the frequency. This signal is the nonlinear error generated by the ADC during the sampling process, as given by the formula... Sure, Represents the amplitude of the second harmonic and This indicates the phase relationship between the second harmonic and the main lobe, and The third harmonic of the signal This is a small signal generated by the main lobe of the signal at three times the frequency. This signal represents the nonlinear error generated by the ADC during the sampling process, as shown by the formula... Sure; Similarly, the kth harmonic of the signal This is a small signal generated by the main lobe of the signal at k times the frequency. This signal is the nonlinear error generated by the ADC during the sampling process, as given by the formula... Sure; Wherein, the normalized frequency f min and f max Signal amplitude and Harmonic order k, second harmonic amplitude and Third harmonic amplitude and And all the way up to the kth harmonic amplitude and All settings are configured as needed.

3. The ADC calibration method based on a fully connected neural network according to claim 2, characterized in that, t m，1 ~∪(0,1000×2πf max )。 4. The ADC calibration method based on a fully connected neural network according to claim 3, characterized in that, The ideal quantization process is described by the formula Q = round(S) m，n The LSB ( / LSB) × LSB is determined, where LSB represents the minimum resolution of the ideal ADC; and the accuracy relationship between the target ADC and the ideal ADC is N. train =N real +2, where N real For the accuracy of the target ADC, N train This represents the accuracy of an ideal ADC.

5. The ADC calibration method based on a fully connected neural network according to claim 4, characterized in that, In the neural network, the weights of the first layer are set to have a mean of 0 and a variance of 0. The values ​​follow a normal distribution with all biases equal to 0. The weights of the second layer satisfy the following conditions: mean 0 and variance 0. The data follows a normal distribution with all biases set to 0. The activation function is ReLU, the loss function is MSE, the optimization method is Adam, and the initial learning rate is 0.1.

Images