Neural network interpretation of aeromagnetic data

• US 5,355,313 A
• Filed: 10/25/1991
• Issued: 10/11/1994
• Est. Priority Date: 10/25/1991
• Status: Expired due to Fees

• Current Assignee: Texaco Incorporated (Chevron Corp.)
• Original Assignee: Texaco Incorporated (Chevron Corp.)
• Inventors: Rogers, John R., Wiener, Jacky M., Pearson, William C., Moll, Robert F.
• Primary Examiner(s): Envall, Jr., Roy N.
• Assistant Examiner(s): CHUNG TRANS, XUONG MY
• Application Number:
  US07/783,016
• Time in Patent Office: 1,082 Days
• Field of Search: 364/420, 364/421, 364/422, 395/23, 395/929
• US Class Current: 702/2
• CPC Class Codes: G01V 3/165 operating with magnetic or electric fields produced or modified by the object or by the detecting device with electromagnetic waves G01V3/17
• CPC Class Codes: G01V 3/38 Processing data, e.g. for analysis, for interpretation, for correction

Abstract

A method for determining depth to basement from aeromagnetic data utilizes neural networks to automate the laborious process of profile interpretation. The neural networks provide consistency, accuracy and overall quality without bias of interpretation.

Claims

1. A computer implemented method for studying structural anomalies along aeromagnetic traces comprising the steps of:

providing a neural network having at least input, intermediate and output layers;

forming at least one two-dimensional profile magnetic basement structural model, each said at least one model having a range of geometric dimension approximating an area to be studied;

reprocessing aeromagnetic survey profiles to remove erroneous readings and the International Geomagnetic Reference Field;

windowing said profiles for input into the network;

training the neural network through weight adjustment using computer generated magnetic profiles from basement structures representative of those anticipated to be found in the basin of interest;

testing the trained network on aeromagnetic profiles where basement structures are already known from other means to validate the network; and

applying the network to recorded aeromagnetic profiles over an unknown portion of the basin.

2. The computer implemented method according to claim 1 further comprising the step of adjusting node weights during the learning process to minimize the error between network prediction and known models according to the Delta Rule of learning and back propagation as understood in the general art of neural networks.

3. The computer implemented method according to claim 1 wherein said two-dimensional profile magnetic basement structural model is computer generated.

4. The computer implemented method according to claim 1 wherein said windowing is accomplished by:

lagging the profile with samples spaced symmetrically to make a window of values.

5. The computer implemented method according to claim 1 further comprising the step of:

creating a cumulative back propagation network of a plurality of processing elements in the input layer, a plurality of processing elements in at least one hidden layer, and at least one element in an output layer.

6. The computer implemented method according to claim 1 further comprising:

affixing a non-linear transfer function to the processing elements in order to achieve optimum learning through weight adjustment.

7. The computer implemented method according to claim 6 wherein said non-linear transfer function is a hyperbolic tangent.

8. The computer implemented method according to claim 1 further comprising:

affixing a linear transfer function to the processing elements in order to achieve optimum learning through weight adjustment.

9. The computer implemented method according to claim 1 further comprising:

allowing the network to learn for a plurality of iterations, then jogging the weighing and resume learning for at least twice the number of iterations.

10. The computer implemented method according to claim 1 further comprising:

testing the network by calculating structural values from the input data.

11. The computer implemented method according to claim 1 used for basement depth determination.

12. The computer implemented method according to claim 1 used to locate basement structures.

13. The computer implemented method according to claim 1 used to locate basement faults.

14. The computer implemented method according to claim 1 used to determine vertical relief of basement structures.

15. The computer implemented method according to claim 1 used to determine separation of structural and compositional anomalies on aeromagnetic profiles.

16. The computer implemented method according to claim 1 used for aeromagnetic profile interpretation for basement mapping.

17. The computer implemented method according to claim 1 further comprising the step of:

creating a cumulative back propagation network of at least one processing element in the input layer, at least one element in at least one hidden layer, and at least one element in an output layer;

affixing a transfer function to the processing elements in order to achieve optimum learning; and

allowing the network to learn for a plurality of iterations, then jogging the weighting and resuming learning for at least twice the number of iterations.

18. The computer implemented method according to claim 17 wherein a processing element can have many input paths coming into it an whose value is the sum of the incoming neuron values times their respective connection weights:

space="preserve" listing-type="equation">I.sub.j Σ

W.sub.ji X.sub.i

where I_j is termed the Internal Activation of the j^th processing element.

19. The computer implemented method according to claim 18 wherein the interval activation of this processing element is modified by a linear transfer function which then becomes input to the next layer of processing elements.

20. The computer implemented method according to claim 17 wherein the matrix of weights are determined or adjusted in response to an input and is referred to as the learning rule and, back-propogation, the responsibility for reducing output error is shared among all of the connection weights, said error first being calculated at the output layer by the difference between the desired output and the actual output, then multiplying this difference by the derivative of the transfer function:

space="preserve" listing-type="equation">E.sub.k.sup.(out) =F'"'"'(I.sub.k.sup.(out))*(D.sub.k -O.sub.k)

and using a sum of products to the previous layer, the previous PE'"'"'s are then assigned proportional error computed in a fashion similar to a feed forward sum of products;

space="preserve" listing-type="equation">E.sub.j.sup.(s) =F'"'"'(I.sub.j.sup.(s))* Σ

.sub.k (E.sub.k.sup.(s+1) *W.sub.kj.sup.(s+1))

21. The computer implemented method according to claim 20 wherein the recursive process back-propagates the error throughout the network and finally the weights are adjusted with the following formula:

space="preserve" listing-type="equation">Δ

W.sub.ji =lcoef*E.sub.j.sup.(s) *X.sub.i.sup.(s-1) E

where;

lcoef=learning coefficient.