Roberto Carluccio_ Rodolfo Conso

Document Sample
Roberto Carluccio_ Rodolfo Conso Powered By Docstoc
					                                                                                                                                                                           Sel1 vs. REB bulletins comparison: looking for underlying features.                                                                                                                                                                                                  SP-12/C

                                                                                                                                                                                Roberto Carluccio, Rodolfo Console, Stefano Chiappini and Massimo Chiappini
                                                                                                                                                                                                            Istituto Nazionale di Geofisica e Vulcanologia, Rome, Italy
The Goal: Improve SEL1 usefulness with a procedure capable of predicting the probability                                                                                                                                                                                                                                                                                                                               GROUP A parameter Matrix of comparisons
that a SEL1 event is not a phase association algorithm artifact

The Idea: mark recent historical SEL1 events as “good” or “bogus” by comparing them with
REB considered here as ground truth and search for hidden patterns among event definition
Procedure: Fit those patterns by means of a neural network trained to classify the new events
as bogus or real.

                                                                                                                                                                                                                                                          In order to find
                                                                                                                                                                                                                                                          a minimum set of
                                                                                                                                                                                                                                                          independent parameters that still
                                                                                                                                                                                                                                                          achieve a good description of our
                                                                                                                                                                                                                                                          events, we have considered two
                                                                                                                                                                                                                                                          groups of parameters:
                                                                                                                                                                                                                                                          (A) directly extracted from the SEL1
                                                                                                                                                                                                                                                          event header and
   Plot of all events stored in DB:                                                                                                                                                                                                                       (B) parameters obtained from the
         blue points: SEL1 events
            red points: REB events                                                                                                                                                                                                                        stations list.

                                                                                                                                                                                                                                                                                                          Both parameters
Full SEL1 and REB bulletins from Jan 2005 to Oct 2008 have been inserted in a database: To limit pro-
                                                                                                                                                                                        For the B group we have tested some geometric parameters obtained from station azimuths and epi-                  groups have been
cessed data amount only a subset of recent bulletin events, from June to Oct 2008 has been conside-
                                                                                                                                                                                        central distances:                                                                                                analyzed by plotting each
red (~20000 events).
                                                                                                                                                                                        • σ(∆ azimuth): standard deviation of differences between consecutive station azimuths.                           of them with respect to any
To every SEL1 event we associate a Boolean “verified flag” that is set (=1) if two conditions are satisfied:
                                                                                                                                                                                        • σ(∆ distance): standard deviation of differences between consecutive epicentral distances.                      other in their group. The resul-
1) the event ID survives in REB (i.e. retains its event_id number)
                                                                                                                                                                                        • Mean[distance]: average epicentral distances                                                                    ting matrixes of plots are shown in
2) the two ID related SEL1-REB events remain close in space and time
                                                                                                                                                                                        • (n * Gap): number of stations times azimuthal Gap                                                               which every considered event is a
As evident in distance and time distributions plot there is in fact a significant fraction of same-ID                                                                                   • p1: sum of the inverse of the epicentral distances to all the operational IMS stations counted                  point colored in green if the SEL1 event is
events that are localized too “far” to be considered “confirmed” .                                                                                                                        positively for stations present in the event stations list, negatively otherwise (see figures below)            confirmed and red otherwise. Specific pat-
                                                                                                                                                                                                                                                                                                          terns show the existence of correlations, the
                                                                                                                                                                                                                                                                                                          evidence of which is promising for their suitable
                                                                                                                                                                                                                                                       GROUP B parameter Matrix of comparisons
                                                                                                                                                                                                                                                                                                          use in the pattern recognition neural network pro-
                                                                                                                        n. of events
             n. of events

                                                                                                                                                                                                                                                                                                          A network supervised training algorithm using “confirmation flag” values has
                                                                  distribution plots of time differences and location geodistances for “same ID” REB-SEL1 events
                                                                                                                                                                                                                                                                                                          been used. In order to optimize network training input a significant, not re-
                                                                                                                                                                                                                                                                                                          dundant subset of input parameters has been looked for with the help of a
                               absolute time difference (sec)                                                                                    Geodetic distance (deg)                                                                                                                                  genetic algorithm search tool. A suitable 12 input subset has been found
                                                                                                                                                                                                                                                                                                          and a network architecture of 12-20-1 has thus been chosen and trained on
We fixed the acceptance thresholds to 5° for the geodetic distance and 30 sec for absolute times dif-                                                                                                                                                                                                     a 15094 records data set.
ferences. The plots below show the discarted (upper) and accepted (lower) ID-related SEL1-REB
events pairs. (pairs connected by thin lines)                                                                                                                                                                                                                                                                                                                                                 Results from test on the Trained Neural Network

                                                                                                                                                                                                  p1 parameter definition
                                                                                                                                                                                                                                                                                                            parameter 1

                                                                                                                                                                                                                                                                                                            parameter 2                                                                                                                         test

                                                                                                                                                                                                                                                                                                                                                           Output “confirmation flag” Layer

                                                                                                                                                                                                                                                                                                                    Input parameter Layer

                                                                                                                                                                                                                                                                                                                                            Hidden Layer
                            REB events
                            discarted SEL1 events
                            REB events related to discarted SEL1 events

                                                                                                                                                                                                                                                                                                           parameter 12

                                                                                                                                                                                                                                                                                                            Feed forward
                                                                                                                                                                                                                                                                                                            12-20-1 neural network

                                                                                                                                                                                            (a)                             (b)

                                                                                                                                                                                                                                                                                                          Different runs of training sequences have been conducted, all showing CCR (Correct Classification
                                                                                                                                                                                                                                                                                                          Rate) values of the order of 75% - 80%. The trained network behavior is shown in term of ROC curve
                                                                                                                                                                                                                                                      a) p1 parameter distributions for flagged (green)   and input-out success-error matrices. On our testing and validating data group, results appear pro-
                                                                                              REB events
                                                                                                                                                                                                                                                      and not flagged (red) events                        mising.
                                                                                              accepted SEL1 events
                                                                                                                                                                                                                                                                               (                 )
                                                                                              REB events related to accepted SEL1 events
                                                                                                                                                                                                                                                      b) distributions ratio
                                                                                                                                                                                                                                                                                   not flagged

Shared By: