1 / 35

Project Goal

Neural Computation Final Project -Earthquake Prediction 0368-4149-01, Spring 2004-5 Alon Talmor Ido Yariv. Project Goal.

julian-rowe
Download Presentation

Project Goal

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Neural ComputationFinal Project -Earthquake Prediction0368-4149-01,Spring 2004-5Alon TalmorIdo Yariv

  2. Project Goal • The task Create an architecture which can provide a tight bound on the variability of the data (when no earthquake is present) and produces increasing abnormalities (outlieres) when earthquake conditions are beginning to develop. An architecture which receives the output of the previous architecture as an input provides a probably score for the development of an earthquake in the next 24 hours.

  3. Earthquake Prediction – the problem • The physics that control earthquakes are at this time poorly understood • Good Sensors are hard to deploy • Earthquakes are very long term – it is hard to collect large amount of statistics

  4. Our Approach • Changes in the eather and Seismic Events are very long term – use long term information. • Good features are extracted by seismologists – use them. • Use a more visual presentation of the data in order to choose the location , features and classifiers to work with.

  5. Hypothesis We Assume two main types of events: • A periodic release of pressure - where the local region of the earth is in a “loop” and releases its pressure periodically. • A sudden release of a large amount of pressure - this event happens after a period of local “silence”. The earth can not release the pressure because of some blocking, pressure is accumulated, and then in one instance a large amount of pressure is released. This usually causes a large earthquake.

  6. Gathering Data Reconstruction of the International Seismological Centre’s Database

  7. International Seismological Centre • A non-governmental organization charged with the final collection, analysis and publication of standard earthquake information from all over the world • The parametric data is available to the general public via various means • Offers a web interface for limited DB queries • Concurrent querying are not allowed, due to bandwidth restrictions (56k/sec)

  8. International Seismological Centre (Cont’d) • The DB holds two kinds of data: Phases and Summery • Phase data - The parametric data collected at each sensor worldwide (Local magnitude, time, etc.) • Summery data - The summery of phase data from all stations, such as the weighted average of the magnitude, the calculated location, time of event, and so on.

  9. International Seismological Centre (Cont’d) • In order to quickly gather the data, we’ve written a Python script • The script queries the DB via the web interface, parses the results, and inserts them into a local database • The local database is based on SQLite, which makes it easy to reuse the data in future research works

  10. ISC’s website

  11. ISC’s Query Page

  12. ISC’s Results Page (Mag.+Phases)

  13. ISC’s Bandwidth Problems

  14. Method – Analyzing the data

  15. Method – Selecting the ROI • We assume events are influnced mostly by activity in close regions. • Region Of Intereset of 3-4 geographical degrees (300-400 Km) is chosen.

  16. Method – Building the basic data • Only MB magnitude and Geographic location are used. • One Dimensional Data is extracted for the chosen ROI. • Data chosen between 2001-2005.

  17. Considering the location - ROI dependant weight matrix

  18. Building the basic data (Cont’d) • DataSample(i) = sqrt(sum(LocationWeightMatrix(Event(j))*Magnitude(Event(j))^2) | 0<j<number of events in current sampling period • The sampling period used in this project is 3600 seconds = 1 hour.

  19. Constructing the features Feature Extracted: • FTT – spectral features • Ceprstum –real(log(PowerFFT(Frame)))

  20. Constructing the classification labels

  21. Feature Dimension reduction Two method used here 1. PCA ( reducing from 20 to ~5 features) 2. The Human Eye

  22. Choosing the model 5 classifiers were compared: • Perceptron • Expectation-Maximization – Gaussian Mixture Models • Radial Basis function • Artificial neural network • Support Vector Machine

  23. Classifying Earthquakes Using SVM

  24. Method • The data was divided (after randomly shuffling it) into two separate sets: • A training set (75%) - Used for constructing the classification model • A testing set (25%) - Used for testing the model • The model was constructed using only the training set, and tested on the testing set

  25. Method (Cont’d) • SVM has several variables which need to be set prior to the model construction process, namely the kernel function and its associated parameters • Since RBF performed well on the data, and it is known as a generally good first pick, we chose it as our kernel function • Thus, there were two free parameters left for us to set

  26. Method (Cont’d) • Cross-validation on the training set was used in order to evaluate various pairs of the parameters • For each pair, the training set was divided into 5 different folds • A SVM model was trained on four folds, and tested on the fifth • This process was repeated for each of the folds (leave one out)

  27. Method (Cont’d) • In order to test the model on a set it is not enough to measure the general accuracy of the model, that is, the number of vectors which the model had successfully labeled • Since there are more negative labeled vectors than positive ones, a trivial classifier which would label any vector as being negative could be evaluated as a relatively good classifier (with a record of over 80% success) Thus, it is important to divide the resulting classification labels into four groups: True Positives, True Negatives, False Positives and False Negatives

  28. Method (Cont’d) • Once statistics of the four above groups was collected, a break-even function was used in order to compare the models: • The parameters of the best model, that is, the model which had the best break-even result were saved, and a model with these parameters was trained on the whole training set • p=True Positives/Sum Positivesn=True Negatives/Sum Negatives

  29. Results • Taking the Sumatra area as a test case (Longitude = 94, Latitude = 8) and labeling any earthquake with a magnitude over 5.5 as positive, the following typical results were achieved:

  30. Results (Cont’d) • Results show that SVM can classify the extracted features quite well, achieving a record of 83.3%/96.28% (True Positive/True Negative) • This is quite surprising, considering the problem's nature • Further research could calibrate the variables more accurately, achieving even better results

  31. The GUI

  32. The GUI (Cont’d)

  33. The GUI (Cont’d)

  34. The GUI (Cont’d)

  35. The GUI (Cont’d)

More Related