{"title": "Lg Depth Estimation and Ripple Fire Characterization Using Artificial Neural Networks", "book": "Advances in Neural Information Processing Systems", "page_first": 544, "page_last": 550, "abstract": null, "full_text": "Lg DEPTH ESTIMATION AND RIPPLE FIRE \n\nCHARACTERIZA TION USING \n\nARTIFICIAL NEURAL NETWORKS \n\nENSCO, Inc. \n\nSignal Analysis and Systems Division \n\nJohn L. Perry and Douglas R. Baumgardt \n\n5400 Port Royal Road \n\nSpringfield, Virginia 22151 \n\n(703) 321-9000, perry@dewey.css.gov \n\nAbstract \n\nThis srudy has demonstrated how artificial neural networks (ANNs) can \nbe used to characterize seismic sources using high-frequency regional \nseismic data. We have taken the novel approach of using ANNs as a \nresearch tool for obtaining seismic source information, specifically \ndepth of focus for earthquakes and ripple-fire characteristics for \neconomic blasts, rather than as just a feature classifier between \nearthquake and explosion populations. Overall, we have found that \nANNs have potential applications to seismic event characterization and \nidentification, beyond just as a feature classifier. In future studies, these \ntechniques should be applied to actual data of regional seismic events \nrecorded at the new regional seismic arrays. The results of this study \nindicates that an ANN should be evaluated as part of an operational \nseismic event identification system. \n\n1 INTRODUCTION \n\n1.1 NEURAL NET,\\VORKS FOR SEISl\\UC SOURCE ANALYSIS \n\nIn this study, we have explored the application of artificial neural networks (ANNs) for \n-the characterization of seismic sources for the purpose of distinguishing between \nexplosions and earthquakes. ANNs have usually been used as pattern matching \nalgorithms, and recent studies have applied ANNs to standard classification between \nclasses of earthquakes and explosions using wavefonn features (Dowla, et al, 1989), \n(Dysart and Pulli, 1990). However, in considering the current state-of-the-art in seismic \nevent identification, we believe the most challenging problem is not to develop a superior \nclassification method, but rather, to have a better understanding of the physics of seismic \nsource and regional signal propagation. \n\n544 \n\n\fLg Depth Estimation and Ripple Fire Characterization \n\n545 \n\nOur approach to the problem has been to use ANN technology as a research tool for \nobtaining a better understanding of the phenomenology behind regional discrimination, \nwith emphasis on high-frequency regional array data, as well as using ANNs as a pattern \nclassifier. We have explored two applications of ANNs to seismic source \ncharacterization: (1) the use of ANNs for depth characterization and (2) the recognition of \nripple-fIring effects in economic explosions. \n\nIn the fIrst study, we explored the possible use of the Lg cross-coherence matrix, \nmeasured at a regional array, as a \"hidden discriminant\" for event depth of focus. In the \nsecond study, we experimented with applying ANNs to the recognition of ripple-fIre \neffects in the spectra of regional phases. Moreover, we also investigated how a small \n(around 5 Kt yield) possibly decoupled nuclear explosion, detonated as part of a ripple-fIre \nsequence, would affect the spectral modulations observed at regional distances and how \nthese effects could be identified by the ANN. \n\n1.2 ANN DESCRIPTION \n\nMLP Architecture: The ANN that we used was a multilayer perceptron (MLP) \narchitecture with a backpropagation training algorithm (Rumelhart, et al, 1986). The \ninput layer is fully connected to the hidden layer, which is fully connected to the output \nlayer. There are no connections within an individual layer. Each node communicates \nwith another node through a weighted connection. Associated with each connection is a \nweight connecting input node to hidden node, and a weight connecting hidden node to \noutput node. The output of \"activation level\" of a particular node is defined as the linear \nweighted sum of all its inputs. For an MLP, a sigmoidal transformation is applied to \nthis weighted sum. Two layers of our network have activation levels. \n\nMLP Training: The:w..,p uses a backpropagation training algorithm which employs \nan iterating process where an output error signal is propagated back through the network \nand used to modify weight values. Training involves presenting sweeps of input patterns \nto the network and backpropagating the error until it is minimized. It is the weight \nthe \nvalues \nrecognition/classification phase. \n\ntrained network and which can be used in \n\nthat represent a \n\nMLP Recognition: Recognition, on the other hand, involves presenting a pattern to \na trained network and propagating node activation levels uni-directionally from the input \nlayer, through the hidden layer(s), to the output layer, and then selecting the class \ncorresponding to the highest output (activation) signal. \n\n2 Lg DEPTH ESTIMATION \n\nIn theory, the Lg phase, which is often the largest regional phase on the seismogram, \nshould provide depth information because Lg results from the superposition of numerous \nnormal modes in the crust, whose excitation is highly depth dependent. Some studies \nhave shown that Lg amplitudes do depend on depth (Der and Baumgardt, 1989). However, \nthe precise dependency of Lg amplitude on depth has been hard to establish because other \neffects in the crustal model, such as anelastic attenuation, can also affect the Lg wave \namplitude. \n\n\f546 \n\nPerry and Baumgardt \n\nIn this study, we have considered if the Lg coherence, measured across a regional array, \nmight show depth dependency. This idea is based on the fact that alI the normal modes \nwhich comprise Lg propagate at different phase velocities. For multilayered media, the \nnormal modes will have frequency-dependent phase velocities because of dispersion. Our \nmethod for studying this dependency is a neural network implementation of a technique, \ncalled matchedjieldprocessing, which has been used in underwater acoustics for source \nwater-depth estimation (Bucker, 1976), (Baggeroer, et al, 1988). This method consists of \ncomputing the spectral matrix of an emitted signal, in our case, Lg, and comparing it \nagainst the same spectral matrix for master events at different depths. In the past, various \noptimal methods have been developed for the matching process. In our study, we have \ninvestigated using a neural network to accomplish the matching. \n\n2.1 SPECTRAL MATRIX CALCULATION AND MATCHED FIELD \n\nPROCESSING \n\nThe following is a description of how the spectral matrix is computed. First, the \nsynthetic seismograms for each of the nine elements of the hypothetical array are Fourier \ntransformed in some time window. If Si (co) is the Fourier transfonn of a time window \nfor the i the channel, then, the spectral matrix is written as, Hij (co) =S j (co) S j*(co), where \nSi(co)=Ale 1[411 +41/ (41)), the indexjis the complex number, rpt is the phase angle, and \nthe * represents complex transpose. The elements, aik of the spectral matrix can be \nwhere the exponential phase shift term \nwritten as a(k co)=AjAie \n\n.j [41, .(11) \n\n\u2022 \n\n( \n\nt1>i (co) -