Research paperParallel optimization of signal detection in active magnetospheric signal injection experiments
Introduction
We introduce new parallel optimization techniques for the detection of signals injected into the Earth's magnetosphere. The application addresses a variety of interesting questions, such as improving an understanding of radiation-belt dynamics, monitoring space weather conditions, and advancing scientific insight into the coupling dynamics between the Earth and the Sun.
Signal detection and extraction is challenging for applications where signal transmission occurs in noisy channels with interference. This is further complicated in this application by interactions of interest between energetic particles and whistler-mode waves in the radiation belts that modify wave behavior and particle populations (Harid et al., 2014a, Harid et al., 2014b, Streltsov et al., 2010, Li et al., 2015). These waves are generated on Earth naturally, for instance by lightning, or artificially, such as by power lines. We examine data from a controlled wave-particle interaction experiment at Siple Station, Antarctica, where a dedicated transmitter injected signals in the extremely and very low frequency ranges (ELF/VLF; 0.3–30 kHz) into the magnetosphere. Such signals travel along field-aligned paths through the magnetosphere, undergoing modifications through the interactions with energetic electrons in the radiation belts, and are received at the geomagnetic conjugate point in Québec, Canada (Liet al, 2014, Helliwell and Katsufrakis, 1974). This experimental setup is shown in Fig. 1.
The received signal is broadband in nature, which requires applying a configurable signal processing chain in order to extract the signal of interest, including for example a lightning filter, a demodulation stage, a smoothing filter, and a signal thresholding filter. The challenge is: how do we select the respective parameters in each stage in order to separate the noise and enhance the detection of the transmitted signal? This problem can lead to large search spaces which can make an exhaustive search intractable.
While the typical signal detection approach uses a manually tuned signal processing chain, alternate parameters from the broader permissible range could provide other insights and emphasize different salient features in the signal. In this work, we automate this search and pruning process while leveraging parallel computing to improve algorithm response time. These performance gains can lead to better scientific insights by exploring more of the parameter space, or enable real-time signal detection. We develop a parallel, multithreaded implementation that achieves good scalability on a multicore platform which improves search throughput and present the results from our evaluation on data from the Siple Station Experiment. Our methods show possible applications for improving the detection of events exhibiting magnetospheric amplification and for highlighting different salient features based on the searched parameter space.
Section snippets
Overview
Siple Station, Antarctica, operated a powerful ELF/VLF transmitter for controlled wave-injection experiments from 1973 to 1988. It enabled a unique experiment in magnetospheric physics, whose observations remain unmatched by any other instrument in providing long-running observations of wave-particle interactions in the magnetosphere (Li, 2015). These data have improved our understanding of various magnetospheric phenomena and processes, but due to their historicity and format, remain an
Problem statement
Prior work (Liet al, 2014, Li et al., 2015) has manually selected parameters for processing data like the one used here in the Siple Station Experiment and then visually detected the transmitted signal. The above mentioned manual process has significant drawbacks and is thus the motivation for this work. We advance methods for the efficient automation of signal detection for the magnetospheric signal injection experiment.
Due to the above mentioned drawbacks of manual signal processing and
Breadth first search and pruning
A useful property of time-series data is that its correlation with a reference signal can be incrementally evaluated after a pipeline stage has been processed. This information can be leveraged to make pruning decisions, i.e., whether to continue applying additional steps in a particular pipeline variant or if the result so far does not warrant further parameter exploration.
In Fig. 5, the output of the stages , , or can be correlated with the reference signal described in Section 2.2.3.
Datasets
The evaluation is conducted using the transmissions from the Siple Station Experiment. The Siple-class consists of two datasets, one containing a signal with an injected transmission (SipleSignal) and the other only containing noise (SipleNoise). To demonstrate that our signal detection method holds across different transmission conditions, we also conduct an evaluation using the Synthetic-class, which consists of a synthetic signal (SyntheticSignal) and a noise (SyntheticNoise) dataset, which
Performance modeling
We advance performance models for the breadth- and depth-first searches in the supplementary material. We accurately predict the response time for different pipeline filtering configurations. For instance, when aggressively filtering, the breadth-first performance model predicts a response time only 2.6 s less than the measured time of 97.07 s.
Related work and discussion
We have used BFS and DFS to optimize the set of parameters for detecting a signal. We used a tree-based representation of the search space that allows for a parallel search strategy, and for data reuse between stages. In this work, using the enumerated search across parameter sets typically yields good parameter values in low-dimensionality search spaces. However, global search strategies could be employed (Bergstra and Bengio, 2012, Powell, 1994).
Other parameter searches are applicable as
Siple detection results
We focused on a single MDIAG transmission at 6/23/1986 7:01:00 UT, SipleSignal, and a single noise dataset at 7:01:15 UT, SipleNoise. We apply the parallel processing parameter search to the 50-min experiment with 100 transmissions of the MDIAG format, from 7:00:00 UT to 7:50:00 UT. Each transmission (at 0 s and 30 s) and noise interval (at 15 s and 45 s) are datasets similar to SipleSignal and SipleNoise. Fig. 10, Fig. 11 show that the performance of our pruning approaches is independent of
Conclusion
This work advances novel automated detection of narrowband signals in the Siple Station Experiment. We improve the detection of signals transmitted through a noisy environment with complex physical interactions by attaching an automated search process to the respective parameterized signal processing workflows.
In addition to improving the effectiveness and scalability of this search, our approach enables reuse of intermediate evaluations among computations exploring alternative parameter values
Acknowledgment
We acknowledge support from NSF ACI-1442997. We thank Frank Lind for useful comments and suggestions on the manuscript.
References (22)
- et al.
Computer aided detection of transient inflation events at Alaskan volcanoes using GPS measurements from 2005-2015
J. Volcanol. Geoth. Res.
(2016) - et al.
Parameter optimization by random search using hybrid computer techniques
- et al.
Random search for hyper-parameter optimization
J. Mach. Learn. Res.
(2012) - et al.
Classification and Regression Trees
(1984) - et al.
Theoretical and numerical analysis of radiation belt electron precipitation by coherent whistler mode waves
J. Geophys. Res.: Space Physics
(2014) - et al.
Finite difference modeling of coherent wave amplification in the Earth's radiation belts
Geophys. Res. Lett.
(2014) - et al.
VLF wave injection into the magnetosphere from Siple Station, Antarctica
J. Geophys. Res.
(1974) - et al.
Time scales of the European surface air temperature variability: the role of the 7–8 year cycle
Geophys. Res. Lett.
(2016) Convergence and efficiency of subgradient methods for quasiconvex minimization
Math. Program.
(2001)- et al.
A Work-efficient Parallel Breadth-first Search Algorithm (Or How to Cope with the Nondeterminism of Reducers)
(2010)