Initial Wave-Type Identification with Neural Networks and its Contribution to Automated Processing in IMS Version 3.0.

摘要

This report describes a new 4-class neural network for automated identification of initial wave type (Teleseism, Regional P, Regional S, or Noise) for data recorded by 3-component stations or arrays. This is an extension of the 2-class (P or S) neural network that we developed for 3-component stations Patnaik and Sereno, 1991. The input data are dominant period, polarization attributes, contextual information (e.g., measurements related to a group of arrivals), a spectral representation of the horizontal-to-vertical power ratio, and the slowness determined by f-k analysis for array stations. We used a three-staged approach, and each stage consists of a 2-class neural network. The first stage separates signal from noise. The signals are passed to the second stage which separates regional S phases from regional P phases and teleseisms. The regional P phases and teleseisms are passed to the final stage which separates them into two distinct classes. A three-layer backpropagation neural network is used at each stage. Neural networks were trained for six 3-component IRIS/IDA stations in the CIS, and a 4-element micro-array in Kislovodsk. The identification accuracy of the neural networks is >90% for most of the stations that we tested. The neural network module was integrated into the Intelligent Monitoring System (IMS), and it was applied to the 3-component IRIS/IDA data under simulated operational conditions. The result was a reduction in the number false-alarms produced by the automated processing and interpretation system by about 60%.