Data processing: measuring – calibrating – or testing – Measurement system in a specific environment – Earth science
Reexamination Certificate
2002-03-08
2004-04-20
Barlow, John (Department: 2862)
Data processing: measuring, calibrating, or testing
Measurement system in a specific environment
Earth science
C702S014000
Reexamination Certificate
active
06725163
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The invention relates to a method for processing a 3-D measurement data set provided with seismic attributes, by means of a neural network, whereby a self-organizing map is trained with selected training data, and the data to be examined are classified according to the self-organizing map.
2. The Prior Art
Neural nets are a general term for a multitude of methods, which follow cognitive learning processes. A discrimination of these methods is based on the type of the learning. In so-called “supervised learning”, both the input and output are prescribed. In the network type that is relevant to this invention, i.e. the self-organizing map, the learning process is unsupervised, and serves to establish relations within the data, and to exploit them. Self-organizing maps originate from the attempt to develop explanations for neural processes in the brain. A stochastic model of that type is described, for example, in the publication Kohonen, T T., 1984, Self-Organisation and Associative Memory, 1. edition, Springer Verlag Heidelberg. It is further referred to the publication Ritter, H., Martinetz, T., and Schulten, K., 1991, Neuronale Netze (Neural Nets), Addison-Wesley.
In the model by Kohonen, all neurons are fully connected to the input. The intensity of the coupling to the input is variable, and is called the weight. Among themselves, the neurons are in each case coupled to their neighbors. If a training pattern is applied, all neurons enter into competition. The neuron specified by the largest similarity with the training pattern wins and adapts its weight towards the pattern. The similarity is computed, for example, by evaluating the scalar product between the training pattern and the weight vector. All other neurons adapt their weights with less strength. In this context, the adaptation strength depends on the distance of the individual neuron to the winner neuron on the neural map. The distance is determined by the number of neurons that are located between the neuron and the winner neuron. Now, further training patterns are applied, and the adaptation step is repeated for each pattern. In this process, the neurons organize themselves towards the training patterns. At the same time, the internal coupling causes all neurons to be involved in the process, and similar training patterns to be imaged to spatially neighbouring neurons.
After the training, the input patterns can be related to the neurons. In this respect, a winner neuron is determined for each training pattern. Desired characteristics of the training patterns can be assigned to the neurons as reference values. In the normal case, the number of input patterns is larger than the number of the neurons, so that several input patterns are imaged to one neuron. The reference values are then formed by the synthesis of the corresponding input patterns. A classification of unknown input data again takes place on the basis of the principle of competition. The pattern is applied to the net, and the reference pattern with the largest accordance is selected. Characteristics of the reference pattern, if existing, can then be transferred to the input pattern.
This generally known method is as well suited as a specialized evaluation method for the complex signal relations in seismic data, in which noisy, non-linear signals have to be processed. In the publication Trappe, H. 1994, Potential neuronaler Netze in der Kohlenwasserstoffexploration und -produktion (Potential of neural nets in the exploration and production of hydrocarbons), conference volume of the 14
th
Mintrop Seminar, the application potential of neural nets for the field of exploration and production of hydrocarbons is presented. In this publication, an example of a self-organizing neural process for the seismic reservoir characterisation is given as well.
From the publication Trappe, H., and Hellmich, C., “Areal Prediction of Porosity Thickness from 3D Seismic Data by Means of Neural Networks” (EAGE 59
th
Conference and Technical Exhibition—Geneva, Switzerland, May 26-30 1977), the application of a neural, self-organizing type of a network for the prediction of the local reservoir quality from 3-D seismic data is known. The aim of the investigation is to establish from seismic data, or from attributes derived from seismic data, a map of porosity thicknesses along an interpreted horizon, here along the Rotliegendes. The amplitude, the lateral variation of the amplitude, and the acoustic impedance obtained from 3-D seismic inversion are used as input data. It is a disadvantage that the processing is limited to a narrowly limited zone, and that only a constant time, or an interpreted horizon is considered.
From the publication De Groot, P., Krajewski, P., and Bischoff, R., 1988, “Evaluation of Remaining Oil Potential with 3D Seismic Using Neural Networks”, EAGE Meeting and Exhibition, Leipzig, the use of an unsupervised neural network for the classification of the pore-fill from seismic data is known. In this context, short traces segments around the investigated horizon are used for the training, and for the classification.
Moreover, from U.S. Pat. No. 5,373,486, a method for the identification and classification of seismological sources, e.g. for the earthquake forecast, or for the verification of tests of nuclear weapons, is known. In this method, a signal arriving at the seismograph is transformed as a time series into a spectogram, and finally into a phase invariant representation by a two-dimensional Fourier transformation. These processed data are presented to a self-organising neural network.
From U.S. Pat. No. 5,940,777 a method for the automatic recognition of seismic patterns is known, in which single seismic trace segments are to be recognized by a self-organizing neural network, where the one-dimensional neural network (chain) exhibits just as much elements, as different patterns are available.
All previously mentioned, known methods have in common, that only single seismic data points (samples) or short seismic trace segments are used, which always refer to only one single seismic trace. It is a disadvantage that the local environment is not considered in the assignment of information derived from the seismic data.
PCT No. WO 97/39367 describes a method and an apparatus for the seismic signal processing and exploration, in which a seismic 3D volume is subdivided into cells. In the simplest case, these cells are cube-shaped. From the trace segments that are located in a cell and that amount to at least two in a cell, a correlation matrix is formed by sums of the differences between inner and outer products of the sets of values from the trace segments. The quotient formed by the highest eigenvalue of the matrix and the sum of all eigenvalues is then calculated as the measure of coherency. As the result, again a 3-D data volume is created, comprised of coherency values. Here sub-volumes are indeed considered, but without application of a neural net.
In the known methods, the training examples needed for the determination of the weights in the neural net, are extracted along a horizon or in a relatively narrow zone around the reservoir of interest. The remaining data is treated in the same way in the classification, which follows the training. The training examples are thus specifically selected, and for this selection, seismic data that was interpreted with other methods before, must be available for the determination of a horizon, or a reservoir. Moreover, the specific selection of training examples implies the risk that physical characteristics of the subsurface which strongly deviate from the selected characteristics, remain without consideration, or at least under-represented, in the classification.
In the state of art, the internal structure and the number of classes of the neural network are furthermore defined before the training. It often turns out, however, that a pre-defined network does not allow a representative classification of the measurement data. Then a multitude of training runs is r
Föll Marc
Hellmich Carsten
Trappe Henning
Barlow John
Collard & Roe P.C.
Henning Trappe
Taylor Victor J.
LandOfFree
Method for processing seismic measured data with a neuronal... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method for processing seismic measured data with a neuronal..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for processing seismic measured data with a neuronal... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3254098