Arrhythmia Detection Using Algorithm and Hardware Co-design for Neural Network Inference Accelerators
使用神经网络推理加速器的算法和硬件协同设计的心律失常检测
The popularization of automatic arrhythmia diagnosis equipment system is helpful to detect the early symptoms of arrhythmia and help people prevent cardiovascular diseases. Nowadays, most of them are machine learning algorithms based on pattern recognition. However, these algorithms have low generalization ability and can't be well applied to a large number of arrhythmia patients. Deep neural network (DNN) has been gradually popularized because of its ability to learn more advanced features from data, showing better generalization ability and robustness. However, DNN/CNN still has some problems that need to be solved urgently, such as the model derivation process consumes a lot of energy and the storage of the model requires a lot of memory space. Therefore, the research of intelligent heart rate detection system with higher energy efficiency has great application prospects in clinical diagnosis, health monitoring and other fields.
Download all versions
git clone https://github.com/SivannaKing/SEU-ASIC-IOT-ECGAI.git
git clone git@github.com:SivannaKing/SEU-ASIC-IOT-ECGAI.git
Download branch or tag of specific version
git clone --branch [branch/tag] https://github.com/SivannaKing/SEU-ASIC-IOT-ECGAI.git
git clone --branch [branch/tag] git@github.com:SivannaKing/SEU-ASIC-IOT-ECGAI.git
-
ver_1_0
基线版本ECG检测AI算法，网络结构为6CONV+2FC，量化使用了分层量化。只有算法部分。
-
ver_1_1
包含算法到硬件实现，并且通过了验证。网络结构为6CONV+1GAP+1FC，量化采用8bit量化。
-
ver_2_0
在ver_1_1的基础上，改进了硬件架构。量化方式采用幂指化。
-
ver_2_1
设计一种基于混合压缩的算法。
-
ver_3_0
二值化版本。
MIT-BIH Arrhythmia Database is available on physionet.
17 classes ECG signals 3600 based on MIT-BIH database is available on mendeley.
5/14 classes ECG signals 3600 based on MIT-BIH database can be generated by MITBIH2ECG_5C14C3600.py according to AAMI EC57 standard.
5 classes QRS signals 130 based on MIT-BIH database can be generated by MITBIH2QRS_5C130.py according to AAMI EC57 standard.
Contributor : 李支青；
基线版本ECG检测AI算法，网络结构为6CONV+2FC，量化使用了分层量化。只有算法部分。
Contributor : 李支青；黄俊光；吴中行；
包含算法到硬件实现，并且通过了验证。网络结构为6CONV+1GAP+1FC，量化采用8bit量化。
- Dataset : MIT-BIH 17 Classes;
- Training Tools : tensorflow; keras;
- Quantize Tools : Qkeras; tensorflow lite;
- RTL Design Tools : Vivado; VCS;
- FPGA ： Xilinx 7020;
Contributor : 黄俊光；
在ver_1_1的基础上，改进了硬件架构。量化方式采用幂指化。
Contributor : 苏峰；
设计一种基于混合压缩的算法。
Contributor : 刘子劲；
二值化版本。5分类。第一层采用8bit。
Contributor : 苗琦慧；
二值化版本。只包含算法。17分类。第一层未二值化。
Contributor : 蒲宁昊；
二值化版本。每一层都二值化了。加速器架构采用流水线的形式。