ECG Classification using Binary CNN on RRAM Crossbar with Nonidealities-Aware Training, Readout Compensation and CWT Preprocessing.

This paper presents an electrocardiogram (ECG) signal classification method using binary CNN implemented on RRAM crossbar arrays. A new RRAM crossbar structure is proposed to provide input-dependent references for adaptive readout quantization. Hence, binary weights can be represented with just one RRAM cell, instead of the conventional differential cell, leading to the reduction of crossbar size by half. Furthermore, the impacts of crossbar nonidealities is mitigated with nonidealities-aware training and in-situ readout compensation. On the other hand, bandpass filter (BPF) based continuous wavelet transform (CWT) approximation is applied for ECG signal preprocessing to enhance the feature extraction. Implemented on 64×64 binary RRAM crossbar arrays, the proposed binary CNN achieved 98.9% classification sensitivity on 10-class MIT-BIH ECG datasets, with only 0.7% sensitivity drop compared to the software version with FP32 weights.