![]() |
|||
![]()
|
![]() |
![]() |
![]() |
2.9. BALDWIN'S SNN FOR MINERAL IDENTIFICATION FROM WELL LOGSBaldwin et al. One of the first efforts that attempted to use simulated neural networks (SNN) in the field of mineral identification and geology is the system that Baldwin et al. built (Baldwin 90) at the end of the 1980s. Once again, well logs and the knowledge-intensive effort of interpretation were the subjects for a pioneering effort. In terms of Baldwin's system, the problem of mineral and lithofacie interpretation from well log data is treated in two steps:
Unlike the original approaches used to tackle the mineral interpretation task, Baldwin's system treats it as a direct vision interpretation task. The effort undertaken to perform the first step was based on the concept of a self-organized network (SON) (Kohonen 88). Self-organizing implies adjusting the network structure as new data becomes available. Hence, self-organization is essentially a self-learning-process. The network that was applied in this case was a fairly sophisticated version assumed to be well suited for vision-oriented recognition tasks. According to Baldwin, visual images can provide information in two ways. The brain and mind can impose a structure on the visual images. If a person thinks about curve shapes or expects a certain shape to be present, he claims information by means of conscious or subconscious willpower. A log analyst that studies different plots might use his interpretation experience and recognition capability to decide that a single, noisy, irregular-shaped data cloud represents two distinct geological features. Later, once this is established, he may view the plots over again and then see the original cloud consisting of two groups of data. In order to handle the pattern identification part of the interpretation task, Baldwin's system applied another type of processing using a different principle that precludes the type of high-level, logic-based operations typical for the other systems presented above. The basic neuron architecture applied for the first step of the procedure was a 2-D self-organizing activation (SOA) concept arranged in an 8D hypercube making room for 88 neurons. Neurons and synapses were represented as elements in an array. Each neuron in one dimension can be characterized by its inhibition and excitation proxies. These are radial limits with respect to deactivation and activation of other neurons. The degree of inhibition and excitation is specific for each neuron and can be described by a mathematical function. The excitation and inhibition qualities of a neuron define the neuron's influence and relative magnitudes. The excitation and inhibition influence changes with time. In order to avoid problems related to available computer power, the resolution of the cube was limited to eight segments per log range. The model ran successfully on an IBM 3090. Once the hypercube organized itself, a final process terminated the first step. This process is called "on-center-off-surrounding" competitive activation. This implies that a set of neurons in a representational group compete among themselves until the strongest neuron eventually surpresses the rest. This allowed a pattern to be determined and consequently decide how many minerals or lithofacies were present. The identified patterns were thus presented before a different network. Because petrophysicists usually recognize log data patterns by deciding among parameter models and, due to the presence of noise and inconsistencies in logs, the so called "Competetive-Activiation Pattern Classification" (CAPC) was applied. This paradigm is described by McClelland (McClelleand 88). The constructed CAPC network consisted of pools of input log neurons and one pool of mineral or lithofacie output neurons. The number of input pools was equal to the number of input logs (see Figure 3). The number of neurons per input log pool equaled the number of segments applied for the log measurements in the previous step (which was 8). Segmentation can be conceived as the resolution along the bore hole axis, as in a geologic column. The discovered number of minerals and lithofacies that the hypercube determined the number of neurons in the output pool. Every neuron in the input pool was connected to every neuron in the output pool. The connections were made bidirectional. The two networks communicate by means of a Hebbian learning rule. This implicates that the neurons in the output pool of the second network were manipulated by means of a training operation.
Baldwin demonstrated that his system was able to apply the hypercube for viewing a complete log interval and discover mineral and lithofacies such as limestone, siltstone, and soft shale. The CAPC network finally generated activations that pinpointed the various discoveries and projected them in the fashion of an ordinary geologic column. An ordinary multivariate, statistically based program was compared to the SNN constructed. Both could identify the prominent lithofacies and minerals. Yet the SNN was able to recognize transient zones between facies and mineralogical units that the other could not. As such, a biologically inspired way of emulating human expertise in a computer environment produced output of value that few humans can match.
|
![]() |
|
Use of this site is subject certain Terms & Conditions. Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details. |