A convolutional neural network (CNN)-based approach for fetal ECG detection from the abdominal ECG recording is proposed. The flow contains a pre-processing phase and a classification phase. In the pre-processing phase, short-time Fourier transform is applied to obtain the spectrogram, which is sent to 2D CNN for classification. The classified results from multiple channels are then fused and high detection accuracy up to 95.2% is achieved and the CNN-based approach outperforms the conventional algorithm. The hardware of this fetal ECG detector composed of the spectrogram processor and 2D CNN classifier is then implemented on the FPGA platform. Because the two dimensions of the spectrogram and the kernel are asymmetric, a pre-fetch mechanism is designed to eliminate the long latency resulted from data buffering for large-size convolution. From the implementation results, it takes 20258 clock cycles for inference and almost 50% computation cycles are reduced. The power consumption is 12.33mW at 324KHz and 1V for real-time operations. The implementation demonstrates the feasibility of real-time applications in wearable devices.

提出了一种基于卷积神经网络 (CNN) 的方法，用于从腹部心电图记录中检测胎儿心电图。该流程包含一个预处理阶段和一个分类阶段。在预处理阶段，应用短时傅立叶变换获得频谱图，然后将其发送到 2D CNN 进行分类。然后融合多个通道的分类结果，实现高达 95.2% 的高检测精度，基于 CNN 的方法优于传统算法。该胎儿心电图检测器的硬件由频谱图处理器和二维 CNN 分类器组成，然后在 FPGA 平台上实现。由于频谱图和内核的两个维度是不对称的，因此设计了预取机制来消除大尺寸卷积数据缓冲导致的长延迟。从实现结果来看，推理耗时20258个时钟周期，减少了近50%的计算周期。在 324KHz 和 1V 下实时操作的功耗为 12.33mW。该实施证明了可穿戴设备中实时应用的可行性。