Download presentation
Presentation is loading. Please wait.
Published byEdward Wilkinson Modified over 9 years ago
1
1 Vertically Integrated Seismic Analysis Stuart Russell Computer Science Division, UC Berkeley Nimar Arora, Erik Sudderth, Nick Hay
2
2 Outline Seismic event monitoring as probabilistic inference Vertically integrated probability models … Connect events to sensor data and everything in between Associate events and detections optimally Automatically take nondetections into account May improve low-amplitude detection and noise rejection Inference using MCMC (poster) Empirical estimation of model components Preliminary experimental results
3
3 Bayesian model-based learning Generative approach P (world) describes prior over what is (source), also over model parameters, structure P (signal | world) describes sensor model (channel) Given new signal, compute P(world | signal) ~ P (signal | world) P (world) Learning Adapt model parameters or structure to improve fit Operates continuously as data are acquired and analyzed Substantial recent advances in modeling capabilities, general-purpose inference algorithms
4
4 Generative model for IDC arrival data Events occur in time and space with magnitude Natural spatial distribution a mixture of Fisher-Binghams Man-made spatial distribution uniform Time distribution Poisson with given spatial intensity Magnitude distribution Gutenberg-Richter Aftershock distribution (not yet implemented) Travel time according to IASPEI91 model+corrections Detection depends on magnitude, distance, station* Detected azimuth, slowness w/ empirical residuals False detections with station-dependent distribution
5
5 Seismic event Travel times Seismic event Travel times Station 1 picks Station 2 picks Generative structure Detected at Station 1? Detected at Station 2? Station 1 noise Station 2 noise
6
6 Inference MCMC (Markov chain Monte Carlo) (see poster S31B- 1713 for details) Efficient sampling of hypothetical worlds (events, travel times, detections, noise, etc.) Converges to true posterior given evidence Key point: computing posterior probabilities takes the algorithm off the table; to get better answers, either Improve the model, or Add more sensors
7
7 Vertical integration: Detection Basic idea: analyzing each signal separately throws away information. Multiple weak signals are mutually reinforcing via a higher- level hypothesis Multiple missing signals indicate that other “detections” may be coincidental noise Simple example: K sensors record either Independent noise drawn from N[0,1] Common signal drawn from N[0,1- ] + independent N[0, ] noise Separate detectors fail completely! Joint detection succeeds w.p. 1 as 0 or K Travel time accuracy affects detection capability!
8
8 STA/LTA Threshold
9
9
10
10
11
11
12
12 Outline Seismic event monitoring as probabilistic inference Vertically integrated probability models … Connect events to sensor data and everything in between Associate events and detections optimally Automatically take nondetections into account May improve low-amplitude detection and noise rejection Inference using MCMC (poster) Empirical estimation of model components Preliminary experimental results
13
13 Seismic event Travel times Seismic event Travel times Station 1 picks Station 2 picks Generative structure Detected at Station 1? Detected at Station 2? Station 1 noise Station 2 noise
14
14
15
15
16
16 Seismic event Travel times Seismic event Travel times Station 1 picks Station 2 picks Generative structure Detected at Station 1? Detected at Station 2? Station 1 noise Station 2 noise
17
17
18
18
19
19 Seismic event Travel times Seismic event Travel times Station 1 picks Station 2 picks Generative structure Detected at Station 1? Detected at Station 2? Station 1 noise Station 2 noise
20
20 Overall Pick Error
21
21 WRA Pick Error
22
22 Overall IASPEI Error
23
23 WRA - IASPEI Error
24
24 Seismic event Travel times Seismic event Travel times Station 1 picks Station 2 picks Generative structure Detected at Station 1? Detected at Station 2? Station 1 noise Station 2 noise
25
25 Overall Azimuth Error
26
26 WRA - Azimuth Error
27
27 Seismic event Travel times Seismic event Travel times Station 1 picks Station 2 picks Generative structure Detected at Station 1? Detected at Station 2? Station 1 noise Station 2 noise
28
28
29
29 Analyzing Performance Min-cost max-cardinality matching where edges exist between prediction and ground truth events within 50 seconds and 5 degrees. Precision – percentage of predictions that match. Recall – percentage of ground truths that match. F1 – harmonic mean of precision and recall. Error – average distance between matching events. (Cost of matching / size of matching)
30
30 Evaluation vs LEB (human experts) F1Precision/ Recall Error/S.D. (km) Average Log- likelihood SEL3 (IDC Automated) 55.646.2 / 69.798 / 119_ VISA (Best Start)80.470.9 / 92.9100 / 117-1784 VISA (SEL3 Start)55.244.3 / 73.4104 / 124-1791 VISA (Back projection Start) 50.649.1 / 52.0126 / 139-1818
31
31 INFERENCE EXAMPLE
32
32 Summary Vertically integrated probability models Connect events, transmission, detection, association Information flows in all directions, reinforcing or rejecting local hypotheses to form a global solution Better travel time model => better signal detection Nondetections automatically play a role Local sensor models calibrated continuously with no need for ground truth May give more reliable detection and localization of lower-magnitude events
33
33 Ongoing Work More sophisticated MCMC design Add more phases and phase relabeling Extend model all the way down to waveforms Evaluation using data from high-density networks (Japan Meteorological Agency, some regions within ISC data)
Similar presentations
© 2024 SlidePlayer.com Inc.
All rights reserved.