一种应用于高精度SAR ADC校准的剪枝神经网络算法

介绍了一种基于剪枝神经网络的后台校准算法,能够对高精度单通道SAR ADC的电容失配、偏移、增益等多个非理想因素同时进行校准,有效提高SAR ADC的精度。本算法不仅可以达到全连接神经网络校准效果,而且同时对贡献小的权重进行剔除,降低了校准电路的资源消耗,加快了神经网络校准算法速度。仿真结果表明,信号频率接近奈奎斯特频率的情况下,对16 bit 5 MS/s的 SAR ADC进行校准,校准后ADC的有效位数从7.4 bit提高到15.6 bit,无杂散动态范围从46.8 dB提高到126.2 dB。

A Pruned Neural Network Algorithm for High Precision SAR ADC Calibration

A background calibration algorithm based on pruning neural network was introduced, which could simultaneously calibrate multiple non-ideal factors such as capacitance mismatch, offset and gain of high-precision single-channel SAR ADC, and effectively improved the accuracy of SAR ADC. This algorithm could not only achieve the full connected neural network calibration effect, but also eliminate the weights with small contributions, which reduced the resource consumption of the calibration circuit and speeded up the neural network calibration algorithm. The simulation results showed that when the signal frequency was close to the Nyquist frequency, the 16 bit 5 MS/s SAR ADC was calibrated, and after calibration, the effective number of bits of the ADC was increased from 7.4 bit to 15.6 bit, and the spurious free dynamic range was increased from 46.8 dB to 126.2 dB.