CN116760412A - An ANN-based time-interleaved ADC calibrator - Google Patents

Abstract

The invention belongs to the technical field of integrated circuits, and particularly relates to an ANN-based time interleaving ADC calibrator. The calibrator provided by the invention comprises two functional networks, a nonlinear calibration network and a time sequence mismatch calibration network, wherein the nonlinear calibration network is used for inhibiting nonlinearity of each channel ADC in the TI ADC, so that nonlinear mismatch among channels is calibrated, and the time sequence mismatch calibration network is responsible for calibrating the time sequence mismatch among the channels. The TI ADC calibrator based on the ANN can calibrate the single-channel nonlinearity, the nonlinear mismatch between channels, the offset mismatch, the gain mismatch and the offset mismatch of the TI ADC at the same time, and improves the overall performance of the TI ADC. Due to the adoption of the ANN technology, the calibrator completes the calibration function after training based on a large amount of data, and has good robustness and universality.

Description

An ANN-based time-interleaved ADC calibrator

Technical field

The invention belongs to the technical field of integrated circuits, and specifically relates to an ANN-based time interleaved ADC calibrator.

Background technique

The principle of a time-interleaved analog-to-digital converter (Time-Interleaved ADC, referred to as TI ADC) is shown in Figure 1. N sub-ADCs are combined into an ADC system to interleave output, thereby increasing the sampling rate of the ADC system by N times. However, time-interleaved ADCs are susceptible to nonlinearity and channel-to-channel mismatch, which introduce spikes and harmonics in the output spectrum, thereby degrading the overall performance of the ADC. Therefore, calibration of various errors in time-interleaved ADCs is essential.

Although calibration techniques for time-interleaved ADCs are now extensively studied, existing methods based on digital models still suffer from issues such as limitations on input signals, additional circuit design, or high-precision references. In addition, calibrations for different errors may affect each other, and the complex interactions between the ADC and various calibration modules make the design and control of the calibration process cumbersome. In view of these limitations, this paper proposes a time-interleaved ADC calibrator based on Artificial Neural Network (ANN), which can calibrate inter-channel nonlinearity and timing mismatch with high accuracy and suppress spurs. and harmonic capabilities superior to previous work. In addition, ANN-based calibrators have also proven to be effective in calibrating random and broadband signals.

ANN is a computational model inspired by the human nervous system and is used to solve complex machine learning and artificial intelligence tasks. ANN consists of a large number of artificial neurons (also called nodes or units) that are connected to simulate the interactions between neurons in the nervous system. ANN is widely used in many fields, including image and speech recognition, natural language processing, predictive analysis, pattern recognition, etc. It has good nonlinear modeling capabilities and adaptability, and can learn and extract features from large amounts of data. With the advancement of artificial intelligence technology, ANN has the potential to be used for error calibration of ADCs to improve performance and simplify the design complexity of the analog front-end. Furthermore, ANNs are known to benefit from data augmentation with noise and other random errors during the training phase, which contributes to increased robustness in the context of ADC calibration.

Contents of the invention

The present invention proposes an ANN-based ADC calibration mechanism, and on this basis, proposes an ANN-based TI ADC calibrator.

The proposed ADC calibration mechanism is shown in Figure 2. The idea is to realize nonlinear vector mapping through ANN, thereby reconstructing the input signal of the ADC and suppressing the harmonics and spurious generated inside the ADC. Treat the analog input of the ADC (also the expected ideal output) as a vector in the time domain (Sequence [x ₁ ,x ₂ ,...,x _L ], where x _k represents an input sample and L represents the vector length), treat the actual output as a vector/> (Sequence [y ₁ , y ₂ ,..., y _L ], where y _k represents an output sample), VRM calibrates the output from the ADC/> Recovery vector/>

Note that the mapping implemented by ANN needs to be bijective to ensure that the input and output of the calibrator are in one-to-one correspondence. In order to ensure that the ANN achieves bijective vector recovery mapping, the present invention proposes two methods: the dimension expansion method and the length increase method. The idea of the dimension expansion method is to expand the ANN by adding additional parameters. The dimensions of , expressed as:

Among them, α _k and β _k are error-related parameters, such as signal derivative values used to calibrate frequency-related errors (such as timing mismatch errors in TI ADC), and used to calibrate encoding-related errors (such as pipelined inter-stage gain error in the ADC) encoding of each stage, etc. By expanding the input vector dimensions, introducing more information, making the ANN more likely to achieve bijection. In addition, more dimensions provide ANN with more degrees of freedom to better represent complex ADC input-output relationships, thereby achieving bijection.

The length increasing method refers to increasing the input vector of the ANN The length L. As L increases, the vector/> It carries more signal information, which makes the mapping through ANN tend to be bijective. In addition, longer input vectors allow for a finer partitioning of the mapped input space, which helps reduce collisions and ambiguities in the map and improves the accuracy and robustness of the ANN. Therefore, ANNs with larger L are more likely to complete bias mapping and recover the expected signal.

On the basis of the above theory, the present invention proposes a TI ADC calibrator based on ANN. The block diagram of the calibrator is shown in Figure 3. The calibrator contains two functional networks, the Nonlinearity Calibration Network (NCN) and the Timing Mismatch Calibration Network (TMCN). NCN is designed to suppress the nonlinearity of each channel ADC in TI ADC, thereby calibrating the nonlinear mismatch between channels. On the other hand, TMCN is responsible for the calibration of timing mismatch between channels. Each channel output of the TI ADC is fed into the NCN for nonlinear calibration, and then offset mismatch and gain mismatch are eliminated through simple addition and multiplication. Then, the calibrated output of each channel is interleaved and combined, and sent to TMCN together with the corresponding derivative value for timing mismatch calibration. The output of TMCN is the expected calibration result and the final output of the entire calibrator.

The specific structures of NCN and TMCN are as follows:

The structure of NCN is shown in Figure 4. It is a convolutional neural network (CNN for short). The network contains two layers, namely the calibration layer and the output layer. The calibration layer includes a calibration module and a matching module, which is used to perform convolution and residual addition processing to obtain features for calibration. The output layer maps features to calibrated output vectors. The error calibration function of NCN is mainly implemented in the calibration layer, in which the calibration module uses ReLU (ReLU(x)=max(0,x)) as the activation function to perform element convolution to complete a short-cut of the residual. The network parameters to achieve the best performance NCN were determined through simulation: the number of feature channels was set to 32, a 1×1 convolution kernel was used in the output layer and matching module, and a 1×3 convolution kernel was used in the calibration module.

TMCN is specifically designed for timing mismatch calibration in TI ADCs. The dimension expansion method is applied to this network to achieve bijective vector reply mapping:

Among them, y _k is an output sample of TI ADC, and y _k ′ represents the first derivative value corresponding to y _k , and N is the number of channels. As shown in Figure 5, TMCN adopts a fully connected neural network structure (fully connected neural network, referred to as FCNN) because it has superior performance and simple structure compared with other networks. The designed network contains three layers, namely input layer, hidden layer and output layer, all of which are fully connected and use ReLU as the activation function. The output results of each layer of FCNN after activation are expressed as:

Where x _j is the j-th input of the layer, and _yi is the output of the i-th neuron, w _ij and _bi are the weight and bias values respectively. The input layer of TMCN contains 2N nodes, the hidden layer contains 2048 nodes, and the output layer contains N nodes. The input of this network is the TI ADC output sequence of length N and the corresponding derivative value, and the output is the calibrated output vector of length N with the timing mismatch eliminated, which is also the final output of the entire calibrator.

The technical solution of the present invention is:

An ANN-based time interleaving ADC calibrator, including a nonlinear calibration module, a mismatch module, an interleaving module, a bandpass differential filter and a timing mismatch calibration module;

The nonlinear calibration module is composed of N nonlinear calibration network units, corresponding to the N output channels of the time interleaved ADC, and defines the output codeword [b0, b1, b2,..., b of the kth output channel of the time interleaved ADC (B-1)]k is converted into quantized value y _k , then y _k is the input of the kth nonlinear calibration network unit, k∈N; each nonlinear calibration network unit consists of the first matching module and the second matching module , the first calibration module, the second calibration module, the third calibration module, the fourth calibration module and the output layer, in which the input data enters the nonlinear calibration network unit and is input into the first matching module and the first calibration module respectively. The first The output data of the calibration module is input to the second calibration module. The output data of the second calibration module is combined with the output data of the first matching module and then input to the second matching module and the third calibration module respectively. The output data of the third calibration module is input To the fourth calibration module, the output data of the fourth calibration module and the output data of the second matching module are combined and input to the output layer, and the output layer outputs the nonlinearly calibrated data of the corresponding channel; the first matching module and the second matching module The module is a convolution layer. The first calibration module, the second calibration module, the third calibration module and the fourth calibration module use ReLU as the activation function to perform element convolution to complete a residual shortcut. In the first matching module, the second calibration module The matching module and the output layer use a 1×1 convolution kernel, while the 1×3 convolution kernel is used in the first calibration module, the second calibration module, the third calibration module and the fourth calibration module;

The mismatch module is used to compensate the offset mismatch and gain mismatch for the nonlinear calibration data output by each nonlinear calibration network unit, and to compensate the nonlinear calibration data of the kth output After mismatch compensation, the vector is obtained/>

Among them, Δo _k is the offset mismatch compensation, Δg _k is the gain mismatch compensation;

The interleaving module is used to perform all the vectors obtained Interleave to obtain the interleaved vector/>

where m represents the round of single-channel conversion;

The bandpass differential filter is used for Process to obtain the corresponding first-order derivative sequence/>

The timing mismatch calibration module is based on and/> Calibrate the timing mismatch Δt _k . The timing mismatch calibration module is a fully connected neural network, including an input layer, a hidden layer and an output layer. ReLU is used as the activation function. The input layer contains 2N nodes, corresponding to/> and/> The output layer contains N nodes. The calculation of each layer of FCNN after activation is expressed as:

where x _j is the j-th input of the layer, _yi is the output of the i-th neuron in the layer, w _ij is the weight of the neuron, and b _i is the bias of the i-th neuron in the layer. by adding and/> Each element in is directly input to the input layer of FCNN. FCNN can complete the calibration of the timing mismatch Δt _k through forward propagation between neurons, corresponding to the vector after completing the timing mismatch compensation. Also the output of the calibrator.

The workflow of the entire calibrator is shown in Figure 6. The specific steps are as follows:

S1. A time-interleaved ADC with a number of bits B quantizes the input signal into a digital signal;

S2. Convert the output codeword [b ₀ , b ₁ , b ₂ ,..., b _(B-1) ] _k of the k-th channel into the corresponding quantized value y _k ;

S3, NCN _k calibrates the output of the k-th channel. The input of NCN _k is the k-channel ADC output vector of length L. The output is a nonlinearly calibrated sequence within a channel of length L/>

S4. Output vector of NCN _k Compensate the traditional offset mismatch _Δok and gain mismatch Δg _k : Compensate by subtracting the offset mismatch _Δok and then dividing by (1+Δg _k ) to obtain the vector/> The specific expression is:

S5, will Interleave to obtain the interleaved vector/> Expressed as

Where N is the number of channels of TI ADC, k represents the kth channel, and m represents the round of single-channel conversion;

S6, will Input to bandpass differential filter to get/> Corresponding derivative sequence/> The impulse response h _bd [n] of the bandpass differential filter is expressed as:

where is the Hilbert transform value corresponding to y[n], It is a rounding down function, k ^NB indicates that the input signal is in the k-th ^NB Nyquist zone of TIADC. And h _d [n] and h _H [n] are the impulse responses of the first-order differentiator and Hilbert transformer respectively, where the L-tap first-order differentiator h _d [n] is expressed as:

And the L-tap Hilbert transformer h _H [n] is expressed as:

S7, will and/> Input to TMCN for calibration of timing mismatch Δt _k . The input of TMCN is the N-length TIADC output and its corresponding derivative value sequence, and the output is an N-length vector after completion of timing mismatch compensation/> Also the output of the entire ANN-based TI ADC calibrator.

S8. The TI ADC calibrator based on ANN completes the calibration.

The beneficial effects of the present invention are:

The ANN-based ADC calibration mechanism proposed by the present invention is a universal and scientific theory, which provides systematic theoretical support for the ANN-based ADC calibration. The ANN-based TI ADC calibrator proposed by the present invention can simultaneously calibrate the single-channel nonlinearity, inter-channel nonlinear mismatch, offset mismatch, gain mismatch and offset mismatch of the TI ADC, thereby improving the overall performance of the TIADC. Due to the use of ANN technology, the calibrator completes the calibration function based on a large amount of data training, and has good robustness and versatility.

Description of the drawings

Figure 1 is a schematic block diagram of a time interleaved ADC.

Figure 2 is a block diagram of the ANN-based ADC calibration mechanism proposed by the present invention.

Figure 3 is a block diagram of the TI ADC calibrator based on ANN proposed by the present invention.

Figure 4 is a schematic structural diagram of the NCN in the ANN-based TI ADC calibrator proposed by the present invention.

Figure 5 is a schematic structural diagram of TMCN in the ANN-based TI ADC calibrator proposed by the present invention.

Figure 6 is a calibration flow chart of the ANN-based TI ADC calibrator proposed by the present invention.

Figure 7 is an effect diagram of simulation verification using the calibrator of the present invention.

Figure 8 is a test verification effect diagram of on-chip integrated calibration using the calibrator of the present invention.

Figure 9 is a test verification effect diagram of off-chip calibration using the calibrator of the present invention.

Detailed ways

The technical solution of the present invention has been described in detail in the content of the invention. The following takes a 12-bit four-channel ADC as an example to illustrate the practicality of the present invention in combination with the drawings and simulation and test examples.

Example

This example is a calibration performance simulation for a 12-bit four-channel 4GSps ADC, N=4, the input signal frequency is 230.68MHz, and the amplitude is -1dBFS. The specific workflow of the calibrator is as follows:

S1. The time-interleaved ADC quantizes the input signal into a digital signal;

S2. Convert the output codeword [b ₀ , b ₁ , b ₂ ,..., b ₁₁ ] _k of the kth (k=0,1,2,3,4) channel into the corresponding normalized quantized value y _k :

y _k ＝2 ¹¹ b ₀ +2 ¹⁰ b ₁ +2 ⁹ b ₂ +...+2 ¹ b ₁₀ +2 ⁰ b ₁₁ (7)

S3. For k=1,2,3,4, NCN _k calibrates the output of the k-th channel. The input of NCN _k is the k-channel ADC output vector of length L=9. The output is the nonlinearly calibrated sequence in the channel of L=9/>

S4. For k=1,2,3,4, the output vector of NCN _k Compensate the traditional offset mismatch _Δok and gain mismatch Δg _k : Compensate by subtracting the offset mismatch _Δok and then dividing by (1+Δg _k ) to obtain the vector/> The specific expression is:

S5. For k=1,2,3,4, the Interleave to obtain the interleaved vector/> Expressed as

Where N is the number of channels of TI ADC, k represents the kth channel, and m represents the round of single-channel conversion;

S6, will Input to a 33-tap first-order bandpass differential filter to obtain/> Corresponding derivative sequence/> Since the input is in the first Nyquist zone, k ^NB = 1 in the first-order bandpass differential filter, the impulse response at this time is,

S7, will and/> Input to TMCN for calibration of timing mismatch Δt _k . The input of TMCN is the TIADC output with a length of 4 and its corresponding derivative value sequence. The output is a vector with a length of 4 after completing the timing mismatch compensation. Also the output of the entire ANN-based TI ADC calibrator.

S8. The TI ADC calibrator based on ANN completes the calibration.

The ANN-based TI ADC calibrator designed by this invention is used for simulation verification in a 12-bit 4GSps four-channel TI ADC model. The model has single-channel interstage gain error, DAC error, noise, high-order harmonics and Offset/gain/timing mismatch between channels. Figure 7 shows the ADC output spectrum before and after using the ANN-based TI ADC calibrator designed by the present invention. It can be seen that the present invention greatly suppresses harmonics and spurs. After calibration, the SNDR and SFDR of the ADC are reduced from 32.20dB to 32.20dB respectively. and 32.76dB increased to 65.35dB and 87.05dB.

In addition, the ANN-based TI ADC calibrator designed by the present invention has been verified on two actual ADC chips: one is a 12-bit 600MSps four-channel ADC prototype chip with an on-chip integrated ANN-based TI ADC calibrator. The other is a 12-bit 5.4GSps four-channel commercial ADC chip that implements an ANN-based TI ADC calibrator off-chip.

Figure 8 shows the test results before and after calibration of the 12-bit 600MSps four-channel ADC prototype chip integrating the ANN-based TI ADC calibrator designed by the present invention in the single tone test. It can be seen that the ANN-based TI ADC calibrator of the present invention is TIADC calibration effectively suppresses harmonics and spurs, improving the SNDR and SFDR of the ADC from 32.79dB and 35.30dB to 62.45dB and 74.21dB respectively.

Figure 9 shows the test results of the 12-bit 5.4GSps four-channel commercial ADC chip before and after calibration in the single-tone test using the ANN-based TI ADC calibrator designed by the present invention for off-chip calibration. It can be seen that the present invention Effectively suppresses the harmonics and spurs of the ADC itself. After ANN-based TI ADC calibration, the SNDR and SFDR of the ADC improved from 42.38dB and 43.17dB to 53.98dB and 78.25dB respectively.

Claims (1)

1. A time-interleaved ADC calibrator based on ANN, characterized by including a nonlinear calibration module, a mismatch module, an interleaving module, a bandpass differential filter and a timing mismatch calibration module; The nonlinear calibration module is composed of N nonlinear calibration network units, corresponding to the N output channels of the time interleaved ADC, and defines the output codeword [b0, b1, b2,..., b of the kth output channel of the time interleaved ADC (B-1)]k is converted into quantized value y _k , then y _k is the input of the kth nonlinear calibration network unit, k∈N; each nonlinear calibration network unit consists of the first matching module and the second matching module , the first calibration module, the second calibration module, the third calibration module, the fourth calibration module and the output layer, in which the input data enters the nonlinear calibration network unit and is input into the first matching module and the first calibration module respectively. The first The output data of the calibration module is input to the second calibration module. The output data of the second calibration module is combined with the output data of the first matching module and then input to the second matching module and the third calibration module respectively. The output data of the third calibration module is input To the fourth calibration module, the output data of the fourth calibration module and the output data of the second matching module are combined and input to the output layer, and the output layer outputs the nonlinearly calibrated data of the corresponding channel; the first matching module and the second matching module The module is a convolution layer. The first calibration module, the second calibration module, the third calibration module and the fourth calibration module use ReLU as the activation function to perform element convolution to complete a residual shortcut. In the first matching module, the second calibration module The matching module and the output layer use a 1×1 convolution kernel, while the 1×3 convolution kernel is used in the first calibration module, the second calibration module, the third calibration module and the fourth calibration module; The mismatch module is used to compensate the offset mismatch and gain mismatch for the nonlinear calibration data output by each nonlinear calibration network unit, and to compensate the nonlinear calibration data of the kth output After mismatch compensation, the vector is obtained/> Among them, Δo _k is the offset mismatch compensation, Δg _k is the gain mismatch compensation; The interleaving module is used to perform all the vectors obtained Interleave to obtain the interleaved vector/> where m represents the round of single-channel conversion; The bandpass differential filter is used for Process to obtain the corresponding first-order derivative sequence/> The timing mismatch calibration module is based on and/> Calibrate the timing mismatch Δt _k . The timing mismatch calibration module is a fully connected neural network, including an input layer, a hidden layer and an output layer. ReLU is used as the activation function. The input layer contains 2N nodes, corresponding to/> and/> The output layer contains N nodes. The calculation of each layer of the fully connected neural network after activation is expressed as: where x _j is the j-th input of the layer, y _i is the output of the i-th neuron in the layer, w _ij is the weight of the neuron, b _i is the bias of the i-th neuron in the layer, by and/> Each element in is directly input to the input layer of the fully connected neural network. The fully connected neural network can complete the calibration of the timing mismatch Δt _k through forward propagation between neurons, corresponding to the vector after completing the timing mismatch compensation/> Also the output of the calibrator.