A novel multi-layer modular approach for real-time fuzzy-identification of gravitational-wave signals

TL;DR

This paper introduces a layered, low-complexity machine learning framework inspired by speech processing for real-time gravitational wave detection, emphasizing modularity and robustness over pure accuracy.

Contribution

The paper presents a novel multi-layer modular framework for gravitational wave detection that is computationally efficient and adaptable, differing from existing complex neural network approaches.

Findings

Achieves 45% detection rate for low SNR signals at 1% false alarm rate.

Lower accuracy compared to state-of-the-art neural networks, but with significantly reduced computational complexity.

Framework's features are robust against the time-position of signals, facilitating real-time applications.

Abstract

Advanced LIGO and Advanced Virgo ground-based interferometers are instruments capable to detect gravitational wave signals exploiting advanced laser interferometry techniques. The underlying data analysis task consists in identifying specific patterns in noisy timeseries, but it is made extremely complex by the incredibly small amplitude of the target signals. In this scenario, the development of effective gravitational wave detection algorithms is crucial. We propose a novel layered framework for real-time detection of gravitational waves inspired by speech processing techniques and, in the present implementation, based on a state-of-the-art machine learning approach involving a hybridization of genetic programming and neural networks. The key aspects of the newly proposed framework are: the well structured, layered approach, and the low computational complexity. The paper describes the basic concepts of the framework and the derivation of the first three layers. Even if the layers are based on models derived using a machine learning approach, the proposed layered structure has a universal nature. Compared to more complex approaches, such as convolutional neural networks, which comprise a parameter set of several tens of MB and were tested exclusively for fixed length data samples, our framework has lower accuracy (e.g., it identifies 45% of low signal-to-noise-ration gravitational wave signals, against 65% of the state-of-the-art, at a false alarm probability of $10^{-2}$), but has a much lower computational complexity and a higher degree of modularity. Furthermore, the exploitation of short-term features makes the results of the new framework virtually independent against time-position of gravitational wave signals, simplifying its future exploitation in real-time multi-layer pipelines for gravitational-wave detection with new generation interferometers.