Research On ADC Nonlinearity Error Calibration Technology Based On Artificial Intelligence Algorithm

As a bridge connecting the analog world and the digital world,Analog-to-Digital Converter(ADC)has a very important reuse in the process of promoting the information industry revolution.Integrated circuit technology has promoted the rapid development of digital chip performance,which also puts forward more requirements for signal acquisition systems,but ADC performance has not improved with the development of semiconductor processes.On the other hand,with the rapid improvement of chip computing power and the hot development of Artificial Intelligence technology,various industries have begun to try to use Artificial Intelligence technology to optimize the performance of products in their field,so ADC calibration algorithm based on artificial intelligence technology has gradually become a new type of calibration algorithm.ADC calibration algorithms based on artificial intelligence have also begun to become a new research direction of calibration algorithms.In this thesis,an ADC calibration algorithm based on Fully Connected Neural Network has been designed.Firstly,the basic principles and performance indicators of ADC are introduced,the structure and main nonlinearity errors of several ADCs are analyzed,and the basic principles of Artificial Intelligence technology and Neural Network learning algorithms are briefly described.Then,based on Python programming,from the perspective of frequency domain,a harmonic-based ADC signal sampling model is established,and the random time series is generated to sample the signal as a data set of the Neural Network,which uniformly contains various characteristics such as amplitude,frequency,and phase of the signal.Then use the Pytorch framework to build a Fully Connected Neural Network model and train and test this network,the initial structure of the network model is set to 30-1200-1,after training,the network model can increase the Effective Number of Bits(ENOB)of the test dataset from 7.5 bits to 12.4 bits.The calibration performance of the network model is further improved by improving the accuracy of the training data set,changing the calibration target to historical signal,and optimizing the network structure,and the ENOB of the calibration test data set is increased from 12.4bit before optimization to13.3bit after optimization,and the structure of the network model becomes 1000-36-1after performance optimization.Using this network model to calibrate the measured data of the actual CT sigma-delta ADC chip and the Pipelined ADC chip,the ENOB of the test data increased by an average of 1.5 bits,and the Spurious Free Dynamic Range(SFDR)increased by an average of 10 d B.Finally,this thesis uses the spatial mapping algorithm based on weight parameters to carry out lightweight and fixed-point research on the neural network model,quantifies the weight of the network from 32-bit floating point to 12-bit fixed-point number through the scaling principle,so that the storage space required by the parameters of the network decreases from the original 141 KB to 53 KB,the calibration performance of the lightweight Neural Network model has almost no change in the ENOB,and the SFDR is only reduced by 3d B.The Artificial Intelligence ADC calibration algorithm architecture proposed in this thesis is suitable for a variety of ADC architectures and can calibrate a variety of nonlinearity errors,providing a new calibration method for the field of ADC calibration algorithms.