Advanced Pulse Signal Processing Algorithm Based on Deep Learning
Title: Advanced Pulse Signal Processing Algorithm Based on Deep Learning
SNIC Project: SNIC 2021/5-55
Project Type: SNIC Medium Compute
Principal Investigator: Shi-Li Zhang <>
Affiliation: Uppsala universitet
Duration: 2021-02-01 – 2022-02-01
Classification: 20299 20205


The aim of our research team is to develop a general and robust signal processing algorithm based on deep learning (specifically ResNet architecture) for feature prediction of noisy pulses found in signals acquired from nanopore sensors and other scenarios. Nanopore sensors possess a wide application scope, including DNA sequencing, protein profiling, nanoparticle characterization, and small chemical molecule detection. Originated from the cell counter, a nanoscale analyte passing through a nanopore immersed in electrolyte will generate a temporary increase of the resistance of the nanopore, due to the volume blockage, which reflects as a pulse in monitoring current traces. To extract information of the analytes from the pulse signals, robust algorithms are urgently demanded to accurately (i) distinguish the pulses from the background noise; and (ii) extract features of the pulses, e.g., amplitude, duration, and appearance frequency. Although several algorithms have been developed, they are based on an empirically user-defined threshold to recognize the pulses. Apparently, the determination of pulses is highly dependent on the how the threshold is specified thereby a risk to become subjective. Furthermore, the subsequent feature extraction is sensitive to the interference of background noise. To resolve the problem, we have proposed an advanced algorithm for feature extraction from pulse signals based on deep learning. The idea has been validated on a small network with small scale datasets and promising results have been obtained. By means of training the network, it is able to acquire the prototype of pulses and possesses the ability of both pulse determination and feature extraction without any a priori assignment of the threshold. It is worth noting that our algorithm is also suitable for any other kinds of pulse signals, e.g., transverse tunneling current based single-molecule sensors, spikes from neural cells and system, electro-cardio pulses, etc. Therefore, our proposed algorithm is a universal, robust, and objective tool for pulse signal processing with an extremely wide application, ranging from biological research, medical inspection, environmental monitoring to the Internet of Things.