COMPUTATIONAL NEURAL NETWORKS FOR GEOPHYSICAL DATA PROCESSING

Transcription

1 SEISMIC EXPLORATION Volume 30 COMPUTATIONAL NEURAL NETWORKS FOR GEOPHYSICAL DATA PROCESSING edited by Mary M. POULTON Department of Mining & Geological Engineering Computational Intelligence & Visualization Lab. The University of Arizona Tucson, AZ USA 2001 PERGAMON An Imprint of Elsevier Science Amsterdam - London - New York - Oxford - Paris - Shannon - Tokyo

2 TABLE OF CONTENTS Preface Contributing Authors xi xiii Part I Introduction to Computational Neural Networks 1 Chapter 1 A Brief History 3 1. Introduction 3 2. Historical Development Mcculloch and Pitts Neuron Hebbian Learning Neurocomputing Perceptron ADALINE Caianiello Neurons Limitations Next Generation 15 Chapter 2 Biological Versus Computational Neural Networks Computational Neural Networks Biological Neural Networks Evolution of the Computational Neural Network 23 Chapter 3 Multi-Layer Perceptrons and Back-Propagation Learning Vocabulary Back-Propagation Parameters Number of Hidden Layers i Number of Hidden Pes Threshold Function Weight Initialization Learning Rate and Momentum Bias Error Accumulation Error Calculation Regularization and Weight Decay Time-Vary ing Data 50 Chapter 4 Design of Training and Testing Sets Introduction Re-Scaling 56

3 3. Data Distribution 4. Size Reduction 5. Data Coding 6. Order of Data Chapter 5 Alternative Architectures and Learning Rules 1. Improving on Back-Propagation 1.1. Delta Bar Delta 1.2. Directed Random Search 1.3. Resilient Back-Propagation 1.4. Conjugate Gradient 1.5. Quasi-Newton Method 1.6. Levenberg-Marquardt 2. Hybrid Networks 2.1. Radial Basis Function Network 2.2. Modular Neural Network 2.3. Probabilistic Neural Network 2.4. Generalized Regression Neural Network 3. Alternative Architectures 3.1. Self Organizing Map Hopfield Networks 3.3. Adaptive Resonance theory Chapter 6 Software and Other Resources 1. Introduction 2. Commercial Software Packages 3. Open Source Software 4. News Groups Part II Seismic Data Processing Chapter 7 Seismic Interpretation and Processing Applications 1. Introduction 2. Waveform Recognition 3. Picking Arrival Times 4. Trace Editing 5. Velocity Analysis 6. Elimination of Multiples 7. Deconvolution 8. Inversion Chapter 8 Rock Mass and Reservoir Characterization 1. Introduction 2. Horizon Tracking and Facies Mans

4 3. Time-Lapse Interpretation Predicting Log Properties Rock/Reservoir Characterization 124 Chapter 9 Identifying Seismic Crew Noise Introduction Current Attenuation Methods Patterns of Crew Noise Interference Pre-Processing Training Set Design and Network Architecture Selection of Interference Training Examples Selection of Signal Training Patterns Testing Analysis of Training and Testing Sensitivity to Class Distribution Sensitivity to Network Architecture Effect of Confidence Level During Overlapping Window Tabulation Effect of NMO Correction Validation Effect on Deconvolution Effect on CMP Stacking Conclusions 153 Chapter 10 Self-Organizing Map (SOM) Network for Tracking 155 Horizons and Classifying Seismic Traces 1. Introduction Self-Organizing Map Network Horizon Tracking Training Set Results \ Classification of the Seismic Traces Window Length and Placement Number of Classes Conclusions 169 Chapter 11 Permeability Estimation with an RBF Network and 171 Levenberg-Marquardt Learning 1. Introduction Relationship Between Seismic and Petrophysical Parameters RBF Network Training Predicting Hydraulic Properties From Seismic Information: Relation 174 Between Velocity and Permeability 3. Parameters That Affect Permeability: Porosity, Grain Size, Clay Content 176

5 4. Neural Network Modeling of Permeability Data Data Analysis and Interpretation Assessing the Relative Importance of Individual Input Attributes Summary and Conclusions 184 Chapter 12 Caianiello Neural Network Method for Geophysical 187 Inverse Problems 1. Introduction Generalized Geophysical Inversion Generalized Geophysical Model Ill-Posedness and Singularity Statistical Strategy Ambiguous Physical Relationship Caianiello Neural Network Method Mcculloch-Pitts Neuron Model Caianiello Neuron Model The Caianiello Neuron-Based Multi-Layer Network Neural Wavelet Estimation Input Signal Reconstruction Nonlinear Factor Optimization Inversion With Simplified Physical Models Simplified Physical Model Joint Impedance Inversion Method Nonlinear Transform Joint Inversion Step 1: MSI and MS Wavelet Extraction At the Wells Joint Inversion Step 2: Initial Impedance Model Estimation Joint Inversion Step 3: Model-Based Impedance Improvement Large-Scale Stratigraphic Constraint Inversion With Empirically-Derived Models Empirically Derived Petrophysical Model for the Trend, Neural Wavelets for Scatter Distribution Joint Inversion Strategy Example Discussions and Conclusions 210 Part III Non-Seismic Applications 217 Chapter 13 Non-Seismic Applications Introduction Well Logging Porosity and Permeability Estimation Lithofacies Mapping Gravity and Magnetics Electromagnetics 225

6 4.1. Frequency-Domain Time-Domain Magnetotelluric Ground Penetrating Radar Resistivity Multi-Sensor Data 230 Chapter 14 Detection of AEM Anomalies Corresponding to 234 Dike Structures 1. Introduction Airborne Electromagnetic Method - Theoretical Background General Forward Modeling for 1 Dimensional Models Forward Modelling for 2 Dimensional Models With EMIGMA Feedforward Computational Neural Networks (CNN) Concept CNNs to Calculate Homogeneous Halfspaces CNN for Detecting 2D Structures Training and Test Vectors Calculation of the Error Term (±lppm, ±2ppm) Calculation of the Random Models (Model Categories 6-8) Training Testing Conclusion 252 Chapter 15 Locating Layer Boundaries with Unfocused 257 Resistivity Tools 1. Introduction Layer Boundary Picking Modular Neural Network Training With Multiple Logging Tools ' Mnn, Mlp, and Rbf Architectures Rprop and Grnn Architectures Analysis of Results Thin Layer Model (Thickness From 0.5 to 2 M) Medium-Thickness Layer Model (Thickness From 1.5 to 4 M) Thick Layer Model (Thickness From 6 to 16 M) Testing the Sensitivity to Resistivity Conclusions 283 Chapter 16 A Neural Network Interpretation System for Near-Surface 286 Geophysics Electromagnetic Ellipticity Soundings 1. Introduction 286

7 2. Function Approximation Background Radial Basis Function Neural Network Neural Network Training Case History Piecewise Half-Space Interpretation Half-Space Interpretations Conclusion 303 Chapter 17 Extracting IP Parameters From TEM Data Introduction Forward Modeling Inverse Modeling With Neural Networks Testing Results Half-Space Layered Ground Polarizable First Layer Polarizable Second Layer Uncertainty Evaluation Sensitivity Evaluation Case Study Conclusions 324 Author Index 327 Index 331