Chapter 4. Neural networks
- Table of Contents
- 4.1. Introduction
- 4.2. Neural network management window
- 4.3. Neural network information
- 4.4. Import GDI networks window
- 4.5. New from PickSets
- 4.6. New from Well Data
- 4.7. NN training window
4.1. Introduction
The neural network module is used to manage, design and train supervised and unsupervised neural networks.
4.1.1. Supervised neural networks
By supervised training, the user is teaching the network to distinguish between two or more pick sets. At each pick location, a number of attribute values is collected.
Just to give an example, fault detection requires two picksets. The user needs to pick locations where a fault is present (pickset A), and locations where there aren't any faults (pickset B). This whole operation can serve two purposes:
Efficiency/automatization: you don't have the time to go through the entire set and pick each and every little fault.
Find 'more than meets the eye': subtle faults can be present without noticing it but the network picks them up anyway.
It will be clear that the interpreter need to define attributes that are sensitive to fault characteristics. Therefore, you need to make picksets of positions that contain different objects ('yes, there is a fault here') and counter-examples ('no, there aren't any faults here'). Other examples include Channel vs. Non-channel or maybe three choices, e.g. Channel deposits, Overbank deposits and Shale. All in all, the very first requirement is that you can pick examples from the data set.
During training of the neural network, you can monitor whether the network can figure out, given the attributes you have defined, whether a location is more like the ones in set A than the ones in set B. If the network would predict that a position would be type A, but in fact it actually is picked by you in pickset B, then that is a misclassification of the network. The lower the number of misclassification, the better.
During application, the software extracts the same attributes at volumes or horizons and uses the same neural network to predict whether we have a type A or type B situation there. When displaying, you can choose between two network outputs:
The probability that a certain position is a 'type A' location
The type number (A=1, B=2 etc.).
At some locations, the network is more 'certain' than at other locations. You can get this 'confidence' by looking at the probability or at the 'Confidence'.
4.1.2. Unsupervised neural networks
In the unsupervised approach you want the network to come up with a 'natural' division of the seismic data. This approach is very useful when you want to perform, for example, horizon-based or volume-based segmentations. After training the network and application of the neural network output to an element, the results should be interpreted.
The (single) pickset holds the example positions at which the software calculates the chosen attributes. Therefore, each position in the pickset will yield a vector of values. The result of the extraction of the attributes at each picked location is the training set.
The neural network tries to cluster this set of vectors. Similar vectors go into the same Segment. This operation can be seen as subdividing the hyperspace of the attribute vectors in compartments. Each compartment has a centre: the cluster centre.
After the training, the network can be applied to a horizon, time-slice or volume. That means that the vectors are extracted in a volume or along a horizon. The network can then classify all those vectors into a Segment. A vector can be close to the cluster centre or further away from it, which is indicated by the Match. The closer to the cluster centre, the higher the match.