An adc calibration method and system based on self-supervised learning neural network

Abstract

This invention relates to the field of integrated circuit analog-to-digital conversion technology, and provides an ADC calibration method and system based on a self-supervised learning neural network. The neural network is constructed, including an encoder and a decoder; the encoder quantizes the high-dimensional ADC result. D error The data is compressed into a low-dimensional latent space, and the decoder reconstructs the low-dimensional latent space data back into the original data space to obtain the calibrated result. D cali The neural network employs a method where the number of neurons in the input and output layers is identical, while the number of neurons in the hidden layers is less than that in either the input or output layers. The quantization results of the ADC at several time points are simultaneously used as training data and labels for the neural network, and training is completed by minimizing the reconstruction error. This method improves the applicability and cost-effectiveness of ADC calibration methods using neural networks.

Description

Technical Field

[0001] This invention relates to the field of integrated circuit analog-to-digital conversion technology, and specifically to an ADC calibration method and system based on a self-supervised learning neural network. Background Technology

[0002] With advancements in integrated circuit technology, analog-to-digital converters (ADCs) have achieved higher speeds and lower power consumption, but also face more severe nonlinearity issues due to smaller device sizes, leading to a decrease in effective resolution. In high-precision applications, calibration algorithms have become crucial for ADC design. Given the complexity of analog calibration, digital calibration has become the mainstream approach. Traditional methods employ PN sequences and LMS algorithms to calibrate the nonlinearity of pipelined ADCs, utilize LMS to calibrate capacitance mismatch in SAR ADCs, or design dedicated calibration schemes for mismatch errors in TI-ADCs. However, these methods are highly dependent on the specific ADC architecture, typically calibrating only a single error source, and their universality and overall performance improvement are limited.

[0003] In recent years, neural networks have been increasingly used in ADC calibration due to their nonlinear mapping capabilities, effectively overcoming the dependence on architecture in traditional methods. For example, Chinese invention application CN118353462A, entitled "A Pipeline ADC Calibration Method Based on GABP Neural Network," generates ADC digital outputs with errors and reference ADC digital outputs by adjusting the error parameters in the ADC model. These two sets of data are used to construct a neural network dataset, and a genetic algorithm is used to optimize the network parameters. The model is then trained using the training set. For example, "High-Speed ​​and Time-Interleaved ADCs Using Additive-Neural-Network-Based Calibration for Nonlinear Amplitude and Phase Distortion" combines dynamic and static networks and shares neurons to reduce overhead; "A Novel NN-Based Fast-Convergence Background Calibration for Timing Mismatch in TI ADCs" constructs a multi-dimensional network to improve performance; "A Neural Network-Enhanced Digital Background Calibration Algorithm for Residue Amplifier Nonlinearity in Pipelined ADCs" utilizes neural networks to extract nonlinear errors; and "Artificial Neural Network Based Calibration for a 12-bit 250 MS / s Pipelined-SAR ADC With Ring Amplifier in 40-nm CMOS" further optimizes hardware through neuron merging. However, the above-mentioned existing methods are all based on supervised learning and rely on high-precision reference ADCs to provide labels, which limits their applicability; moreover, hardware optimization is mostly focused on the structural level, neglecting the optimization potential of the training process. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to improve the applicability and economy of ADC calibration methods using neural networks.

[0005] The present invention solves the above-mentioned technical problems through the following technical means:

[0006] This invention provides an ADC calibration method based on a self-supervised learning neural network. The neural network comprises an encoder and a decoder; the encoder quantizes the high-dimensional ADC result. Derror The data is compressed into a low-dimensional latent space, and the decoder reconstructs the low-dimensional latent space data back into the original data space to obtain the calibrated result. D cali The number of neurons in the input and output layers of a neural network is exactly the same, while the number of neurons in the hidden layers is less than the number of neurons in the input or output layers. The quantization results of the ADC to be calibrated at several time points are used simultaneously as training data and training labels for the neural network, and the network training is completed by minimizing the reconstruction error.

[0007] Furthermore, the encoder is a process from the input layer to the hidden layer, as follows:

[0008] Where W1 and b1 are the weight and bias matrices from the input layer to the hidden layer, respectively. Activ is the activation function; h is the hidden layer output matrix, D error This is the matrix representing the ADC quantization results.

[0009] Furthermore, the decoder is a process from the hidden layer to the output layer, as follows:

[0010] Where W2 and b2 are the weight and bias matrices from the hidden layer to the output layer, h is the hidden layer output matrix, and D... cali This is the calibrated result matrix.

[0011] Furthermore, the number of neurons in the input and output layers of the neural network is greater than the number of points sampled by the ADC in one cycle.

[0012] Furthermore, minimizing the reconstruction error specifically involves using the mean squared error as the loss function for self-supervised learning, as shown in the following equation:

[0013] in, k D represents the number of samples, Derror is the matrix of ADC quantization results, and Dcali is the calibrated result matrix.

[0014] Furthermore, the activation function is the symmetric saturated linear transfer function satlins, as shown in the following equation: .

[0015] This invention also provides an ADC calibration system based on a self-supervised learning neural network. The system executes the above-described method during operation and includes the following modules: The network building block, used to construct neural networks, includes an encoder and a decoder; the encoder quantizes the results of a high-dimensional ADC. D error The data is compressed into a low-dimensional latent space, and the decoder reconstructs the low-dimensional latent space data back into the original data space to obtain the calibrated result. D cali The number of neurons in the input and output layers of a neural network is exactly the same, while the number of neurons in the hidden layers is less than the number of neurons in either the input or output layers. The network training module is used to simultaneously use the quantization results of the ADC to be calibrated at several time points as training data and training labels for the neural network, and completes network training by minimizing the reconstruction error.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method.

[0017] The present invention also provides an adaptive neural network reconstruction method based on ADC frequency. Based on the above-mentioned neural network, by detecting the frequency of the input signal, the neurons in the input layer and output layer of the neural network are adaptively deactivated. Under the premise that the number of neurons in the input layer and output layer of the neural network is greater than the number of points sampled in one cycle of the ADC, the neural network structure is simplified.

[0018] Furthermore, the adaptive inactivation operation on neurons in the input and output layers of the neural network specifically involves: If the initial number of neurons in the input and output layers of a neural network is M, and the oversampling rate N of the input signal is determined by the zero-crossing detection method, then the input and output layers of the neural network are synchronously deactivated with MN neurons, thus simplifying the neural network structure.

[0019] The advantages of this invention are: (1) This invention uses a self-supervised learning neural network, which eliminates the need for a high-precision reference ADC, thus improving applicability while ensuring calibration accuracy; the circuit structure is simple, reducing design difficulty and hardware overhead, and the response speed is fast, making it suitable for high-speed ADC scenarios; at the same time, it is highly economical.

[0020] (2) This invention reduces the number of training iterations by adaptively adjusting the neural network structure and optimizing the number of neurons, thereby accelerating calibration and reducing power consumption. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a conventional ADC calibration algorithm structure based on a multidimensional neural network, as described in an embodiment of the present invention. Figure 2This is an equivalent system block diagram of a conventional ADC calibration algorithm based on a multidimensional neural network, as described in this embodiment of the invention. Figure 3 This is a schematic diagram of the structure of an ADC calibration method based on a self-supervised learning neural network according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the relationship between the number of hidden layer neurons and network performance in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the principle of an adaptive neural network reconstruction method based on ADC frequency according to an embodiment of the present invention. Figure 6 This is a schematic diagram comparing the signal spectra before and after calibration for different frequency input signals in an embodiment of the present invention; Figure 7 This is a schematic diagram comparing circuit consumption and training speed before and after network reconstruction in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 Current neural network-based ADC calibration methods primarily utilize the powerful data fitting capabilities of neural networks. They take the output of the ADC to be calibrated as input and the error between it and the output of a high-precision reference ADC as a label. Through iterative training, a mapping relationship from the target output to the reference output is established. Traditional multidimensional neural network-based ADC calibration structures are as follows: Figure 1 As shown. This algorithm maps input data to output data in a black-box manner. The neural network input layer contains 2... i The hidden layer contains 10 neurons. m 1 neuron. Among them, D error (n) This represents the output of the ADC to be calibrated at multiple adjacent time points. D ref (n) This indicates the output of the high-precision reference ADC. w k,j (L) This represents the weights from the k-th neuron in layer L to the j-th neuron in layer L+1. The loss function compares the network outputs... D cali (n) With reference outputD ref (n) The difference is calculated using the commonly used Mean Squared Error (MSE) function. Weights w k,j (1) ’ and w m,1 (2) ’ The update formula (ignoring the bias term for simplicity) can be expressed as:

[0024]

[0025] in, D hidden (j) It is the output table of the j-th neuron in the hidden layer. Activ This is the activation function. The network continuously adjusts the weights through backpropagation. After training, this parameter is used for forward propagation to obtain the calibrated ADC output. D cali (n)’ As shown in the following formula:

[0026] Since the neural network input is the quantized output of the ADC at multiple adjacent time points, the system block diagram can be equivalently represented as follows: Figure 2 As shown. The system function of the entire calibration system can be expressed as:

[0027] This calibration system can be viewed as a cascade of multiple filters, a nonlinear system, and a linear system. The calibration algorithm is essentially an adaptive filtering system capable of dynamically adjusting parameters based on the input and optimizing tap coefficients through backpropagation. Therefore, this method relies on a high-precision reference ADC, significantly increasing hardware costs; existing research often uses LM fitting as an alternative. Furthermore, such methods have a fixed network structure, leaving room for improvement in training efficiency and resource utilization. Moreover, limited by the network structure design, the algorithm can only calibrate one output per cycle, and the speed limitation of the calibration circuit makes it unsuitable for high-speed ADC scenarios.

[0028] This embodiment explores and mines the internal relationships within ADC data to achieve effective ADC calibration without requiring reference ADC data. It provides an ADC calibration method based on a self-supervised learning neural network, constructing the neural network as follows: Figure 3 As shown, it includes an encoder and a decoder; the encoder quantizes the high-dimensional ADC result. D errorThe data is compressed into a low-dimensional latent space, and the decoder reconstructs the low-dimensional latent space data back into the original data space to obtain the calibrated result. D cali The number of neurons in the input and output layers of the neural network is exactly the same, while the number of neurons in the hidden layers is less than that in either the input or output layer. Traditional neural network-based calibration schemes do not have a fixed number of hidden layers; this number is often determined through testing. In traditional schemes, a higher number of hidden layer neurons generally results in better calibration, but also increases resource consumption. However, the method proposed in this embodiment achieves better calibration results with fewer hidden layers. The role of the hidden layers is to compress the ADC data into a low-dimensional potential space, and their number of neurons must be less than that of the input and output layers. Fewer hidden layer neurons mean less noise and other non-ideal factors are included in the compressed data. This embodiment statistically analyzed the calibration results with different numbers of hidden layers and ultimately set the number of hidden layer neurons to 2. This balances calibration performance with reduced hardware overhead in the calibration circuit.

[0029] The quantization results of the ADC to be calibrated at several time points are used simultaneously as training data and training labels for the neural network, and the network training is completed by minimizing the reconstruction error.

[0030] The encoder is the process from the input layer to the hidden layer, as shown in the following equation:

[0031] Where W1 and b1 are the weight and bias matrices from the input layer to the hidden layer, respectively. Activ is the activation function; h is the hidden layer output matrix, D error This is the matrix representing the ADC quantization results.

[0032] The activation function used is the symmetric saturated linear transfer function, Satlins, which has a linear region and a saturated region. In the linear region ([-1, 1]), its derivative is always 1, which is beneficial for the stable flow of the gradient during backpropagation; in the saturated region (where the absolute value of the input is greater than 1), the output no longer changes, and the derivative is 0, mathematically expressed as follows: .

[0033] The decoder is the process from the hidden layer to the output layer, as shown in the following equation:

[0034] Where W2 and b2 are the weight and bias matrices from the hidden layer to the output layer, h is the hidden layer output matrix, and D... cali This is the calibrated result matrix.

[0035] To ensure that the training data of the neural network contains complete information features, the number of neurons in its input and output layers needs to be greater than the number of points sampled by the ADC in one cycle. Specifically, the number of neurons in the input and output layers should be set as: the fastest sampling frequency of the ADC / the lowest frequency that the ADC can quantize. In this embodiment, based on the typical frequency of the calibration object (10.3MHz) and the fastest sampling frequency (1GHz), the number of neurons in the input and output layers is set to 101.

[0036] The aforementioned minimization of reconstruction error specifically involves using the mean squared error as the loss function for self-supervised learning, as shown in the following equation:

[0037] in, k D represents the number of samples, Derror is the matrix of ADC quantization results, and Dcali is the calibrated result matrix.

[0038] The method described above uses the quantization results of the ADC to be calibrated as label data and trains the network by minimizing the reconstruction error (i.e., the difference between the original input and the decoder output). The key to achieving effective calibration lies in the special setting of low-dimensional hidden layers. To ensure that the decoder output can recover the original input as perfectly or approximately as possible, the network is forced to retain the most critical information in the latent space, thereby learning the underlying patterns and information behind the data. The trained network ultimately becomes a network with denoising and calibration functions.

[0039] The number of neurons in the hidden layer determines the effective data information captured by the network, and thus determines the network's capabilities and performance. For data like that output by an ADC, such as... Figure 4 As shown, the calibration results for the same ADC output are obtained when the number of neurons in the input and output layers is the same (51 in each), but the number of neurons in the hidden layers is different. It can be seen that as the number of hidden layers increases, the calibration effect gradually decreases, and the loss function value also decreases. This is because the hidden space obtained by network dimensionality reduction contains more original signal information. When the number of neurons in each layer tends to be consistent, the network output almost completely replicates the input.

[0040] Example 2 It should be further explained that, based on the same inventive concept, this embodiment also provides an ADC calibration system based on a self-supervised learning neural network. The system executes the method described in Embodiment 1 during operation, including the following modules: The network building block, used to construct neural networks, includes an encoder and a decoder; the encoder quantizes the results of a high-dimensional ADC. D error The data is compressed into a low-dimensional latent space, and the decoder reconstructs the low-dimensional latent space data back into the original data space to obtain the calibrated result. Dcali The number of neurons in the input and output layers of a neural network is exactly the same, while the number of neurons in the hidden layers is less than the number of neurons in either the input or output layers. The network training module is used to simultaneously use the quantization results of the ADC to be calibrated at several time points as training data and training labels for the neural network, and completes network training by minimizing the reconstruction error.

[0041] This embodiment also provides a computer storage medium storing a computer program, which is executed by a processor to perform the method described in embodiment 1.

[0042] For the purposes of this specification, "computer storage medium" can be any means capable of containing, storing, communicating, propagating, or transmitting a program for use in or in conjunction with an instruction execution system, apparatus, or device. More specific examples of computer storage media (a non-exhaustive list) include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, a computer storage medium can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0043] Example 3 This embodiment provides an adaptive neural network reconstruction method based on ADC frequency. Based on the neural network described in Embodiment 1, by detecting the frequency of the input signal, the neurons in the input and output layers of the neural network are adaptively deactivated. Under the premise that the number of neurons in the input and output layers of the neural network is greater than the number of points sampled in one cycle of the ADC, the neural network structure is simplified.

[0044] The adaptive inactivation operation on neurons in the input and output layers of the neural network, such as... Figure 5 As shown, specifically: If the initial number of neurons in the input and output layers of a neural network is M, and the oversampling rate N of the input signal is determined by the zero-crossing detection method, then the input and output layers of the neural network are synchronously deactivated with MN neurons, thus simplifying the neural network structure.

[0045] When dealing with regression problems, neural networks exhibit a strong correlation with the input data. The quantization output of an ADC (Analog-to-Digital Converter) often shows a certain periodicity; therefore, the characteristics of ADC data also exhibit periodicity. Theoretically, a multidimensional neural network only needs to use an integer multiple of the number of sampling points of the ADC within one cycle to ensure that the input dataset covers all data features, thus enabling normal training of the neural network. Figure 5 As shown, the initial size of the neural network is fixed, with M neurons in both the input and output layers. The signal passes through an OSR discriminator, which counts the number of zero-crossings Z within one sampling clock cycle T and calculates the number of sampling points N within one cycle, thereby obtaining the information density of the signal within one cycle. Based on this, the input and output layer structures of the network are reconstructed. The reconstructed network has a more streamlined structure, which greatly reduces the power consumption and resource consumption of the calibration circuit. Furthermore, the network needs to learn less data, and the circuit needs to perform fewer calculations, which greatly accelerates the training and inference speed of the neural network.

[0046] To verify the effectiveness of the proposed algorithm, this embodiment employs an off-chip verification method, using an actual 14-bit 1GSPS pipelined ADC as the test case. The complete test platform includes a 100MHz crystal oscillator and a clock distribution development board to provide a 1GHz sampling clock for the ADC. The neural network is deployed on an FPGA development board (ADS7-V2EBZ), and the ADC and FPGA development board transmit data via the JESD204B protocol.

[0047] The initial architecture of the neural network is set to 101x2x101. For example... Figure 6 As shown in (a), the results before and after calibration are compared when a 10.3MHz sine wave signal is input. The results show that the proposed calibration algorithm can still effectively calibrate the ADC output even without using a reference ADC, with an increase of 3.2 effective bits, and improvements of SNDR and SFDR of 19.6dB and 24.1dB, respectively.

[0048] The structure of the network described above can meet the calibration requirements of a 10.3MHz signal at a 1GHz sampling rate. For a 40.3MHz signal at a 1GHz sampling rate, only a 25x2x25 neural network is needed. By deactivating the neurons in the input and output layers, the results before and after calibration are as follows. Figure 6 As shown in (b), the results show that the network can still effectively complete the calibration task.

[0049] like Figure 7 As shown in (a), the resource overhead and power consumption are compared before and after using the network reconstruction strategy when the input signal is 40.3MHz. The results show that network reconstruction can effectively reduce resource usage and power consumption.

[0050] like Figure 7 As shown in (b), the number of training batches required for a neural network to achieve the same loss value under different network sizes is summarized. The results show that when the number of neurons in the input and output layers is set according to the reconstruction strategy, a lower loss value can be achieved more quickly, thus significantly shortening the training time of the neural network.

[0051] This embodiment also demonstrates a comparison between this method and other methods. Compared with existing methods, this method significantly reduces hardware resource consumption and accelerates the training and construction process of the calibration network while ensuring calibration effectiveness, thus reducing the burden on digital circuits. Details are shown in the table below:

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An ADC calibration method based on a self-supervised learning neural network, characterized in that, Construct a neural network, including an encoder and a decoder; the encoder quantizes the results of a high-dimensional ADC. D error The data is compressed into a low-dimensional latent space, and the decoder reconstructs the low-dimensional latent space data back into the original data space to obtain the calibrated result. D cali The number of neurons in the input and output layers of a neural network is exactly the same, while the number of neurons in the hidden layers is less than the number of neurons in the input or output layers. The quantization results of the ADC to be calibrated at several time points are used simultaneously as training data and training labels for the neural network, and the network training is completed by minimizing the reconstruction error.

2. The ADC calibration method based on a self-supervised learning neural network according to claim 1, characterized in that, The encoder is the process from the input layer to the hidden layer, as shown in the following equation: Where W1 and b1 are the weight and bias matrices from the input layer to the hidden layer, respectively. Activ is the activation function; h is the hidden layer output matrix, D error This is the matrix representing the ADC quantization results.

3. The ADC calibration method based on a self-supervised learning neural network according to claim 1, characterized in that, The decoder is the process from the hidden layer to the output layer, as shown in the following equation: Where W2 and b2 are the weight and bias matrices from the hidden layer to the output layer, h is the hidden layer output matrix, and D... cali This is the calibrated result matrix.

4. The ADC calibration method based on a self-supervised learning neural network according to claim 1, characterized in that, The number of neurons in the input and output layers of the neural network is greater than the number of points sampled by the ADC in one cycle.

5. The ADC calibration method based on a self-supervised learning neural network according to claim 1, characterized in that, The aforementioned minimization of reconstruction error specifically involves using the mean squared error as the loss function for self-supervised learning, as shown in the following equation: in, k D represents the number of samples, Derror is the matrix of ADC quantization results, and Dcali is the calibrated result matrix.

6. The ADC calibration method based on a self-supervised learning neural network according to claim 2, characterized in that, The activation function used is the symmetric saturated linear transfer function, Satlins, as shown in the following equation: 。 7. An ADC calibration system based on a self-supervised learning neural network, characterized in that, Includes the following modules: The network building block, used to construct neural networks, includes an encoder and a decoder; the encoder quantizes the results of a high-dimensional ADC. D error The data is compressed into a low-dimensional latent space, and the decoder reconstructs the low-dimensional latent space data back into the original data space to obtain the calibrated result. D cali The number of neurons in the input and output layers of a neural network is exactly the same, while the number of neurons in the hidden layers is less than the number of neurons in either the input or output layers. The network training module is used to simultaneously use the quantization results of the ADC to be calibrated at several time points as training data and training labels for the neural network, and completes network training by minimizing the reconstruction error.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

9. An adaptive neural network reconstruction method based on ADC frequency, characterized in that, Based on the neural network described in any one of claims 1-6, by detecting the frequency of the input signal, the neurons in the input and output layers of the neural network are adaptively deactivated, thereby simplifying the neural network structure while ensuring that the number of neurons in the input and output layers of the neural network is greater than the number of points sampled by the ADC in one cycle.

10. The adaptive neural network reconstruction method based on ADC frequency according to claim 9, characterized in that, The adaptive inactivation operation of neurons in the input and output layers of the neural network specifically includes: If the initial number of neurons in the input and output layers of a neural network is M, and the oversampling rate N of the input signal is determined by the zero-crossing detection method, then the input and output layers of the neural network are synchronously deactivated with MN neurons, thus simplifying the neural network structure.

Images