Optimizing Autoencoder Training for Efficient Data Compression

The use of heart rate monitoring devices has increased significantly, leading to the emergence of a new class of consumer-grade wearable devices designed for continuous heart rate monitoring. Nevertheless, minimizing the power consumption of these devices has emerged as a significant challenge. In light of this, we propose an intelligent wearable device that employs Bluetooth Low Energy (BLE) technology for the acquisition and transmission of electrocardiogram (ECG) signals. This work aims to design and implement an intelligent wearable device that can continuously acquire and transmit ECG signals using BLE technology, while reducing power consumption, and thereby, enhancing their operational efficiency. The main objective is to reduce power consumption during signal transmission and build an encoder that can encode the signal despite variations in its shape at the source. To achieve this goal, we present an algorithm that uses the Autoencoder (AE) as its main component. The algorithm learns online and adapts to changes in the input source’s shape without discarding previously learned information or expanding memory space. We implemented the algorithm on Raspberry Pi3 and show that the encoder is utilized over 94% of the time, encoding input signal patterns under a specific constraint that the RMSE of the encoded signals should be less than 2%. As a result, we achieved a significant reduction in power consumption, gaining approximately eight times compared to the non-compressed mode.