Optimal design of digital FIR filters based on text convolutional neural network for aliasing errors reduction in FI-DAC

Abstract

This paper presents optimal design of digital finite impulse response (FIR) filters based on text convolutional neural network (TEXT-CNN) to reduce the aliasing errors in frequency interleaving digital-to-analog converters (FI-DAC). For an example of FI-DAC with M sub-DACs, we first obtained the approximation error function. We add the real and imaginary parts of the approximation error to obtain the error function. Finally, the task-relevant spectral features are extracted by the convolution operation of the TEXT-CNN model, and the parameters are updated by iterative training to obtain optimal coefficients of digital FIR filters. Additionally, we derived the computational complexity of our proposed TEXT-CNN architecture. Several design examples were given to verify the performance of our proposed optimal design based on TEXT-CNN. The simulation results showed that, by using our presented optimal design based on TEXT-CNN, the maximum distortion error is 0.0016dB, and the maximum aliasing error is -73.24dB which satisfied the desired spurious free dynamic range (SFDR) in a 12-bit FI-DAC system. Further, the computational complexity of our presented optimal design based on TEXT-CNN is compared with other three optimal designs, our proposed TEXT-CNN optimal design can obtain better aliasing errors reduction at the cost of higher computational complexity.