At the Edge of the Heart: ULP FPGA-Based CNN for On... | AI Research

At the Edge of the Heart: ULP FPGA-Based CNN for On-Device Cardiac Feature Extraction in Smart Health Sensors for Astronauts
As space exploration missions to the Moon and Mars become more frequent, the need for autonomous, reliable health monitoring for astronauts has become a top priority. Because deep-space missions face significant communication delays, astronauts cannot rely on real-time support from ground stations. This paper introduces a specialized, ultra-low-power (ULP) hardware solution that allows wearable sensors to perform real-time cardiac monitoring directly on the device, ensuring that critical health data can be processed and analyzed without needing an external connection.

Monitoring the Heart in Space

The researchers focused on Seismocardiography (SCG), a non-invasive method that uses sensors to detect the mechanical vibrations of the heart. While SCG is effective for monitoring cardiac activity, it is difficult to analyze automatically due to motion artifacts and the physical changes the human body undergoes in microgravity. To solve this, the team developed a Convolutional Neural Network (CNN) capable of identifying specific cardiac phases—systolic and diastolic—directly from raw sensor data.

Efficient Hardware Design

To make this technology suitable for space, the team utilized the Lattice iCE40UP5K, a commercial, ultra-low-power FPGA. Because this hardware is highly resource-constrained, the researchers employed a technique called "Quantization-Aware Training" (QAT). This method trains the neural network to function using 8-bit integers rather than complex floating-point numbers. By combining this with a "systolic-array" accelerator—a grid-like structure that processes data in a highly efficient, pipelined flow—the system can perform complex calculations while consuming very little power and memory.

Performance and Results

The implementation proved to be highly effective for long-duration missions. The system achieved a 98% validation accuracy in classifying cardiac phases while consuming only 8.55 mW of power. It completes the entire inference process in 95.5 milliseconds, using a minimal amount of the FPGA’s internal resources. These findings demonstrate that it is possible to run sophisticated, autonomous health-monitoring AI on small, energy-efficient hardware that is resilient enough to operate in the harsh radiation environments of space.

Key Considerations

The success of this approach relies on the co-design of the neural network model and the hardware architecture. By folding operations like batch normalization into the convolution layers and using a specialized memory subsystem to manage data flow, the researchers bypassed the limitations of the small FPGA. This design strategy ensures that the system remains lightweight and power-efficient, making it a viable candidate for future wearable health sensors in deep-space exploration.