The previous course tells you: In the field of oil and gas reservoir surveying, the geometric forms of faults and fracture networks play an important role in oil and gas accumulation and migration. Therefore, it is necessary and worthwhile to identify them. Fault recognition is so important, how to do it specifically?
01 Manual detection of fault
fault recognition was initially done by experienced interpreters based on the local characteristics of the fault, combined with the geological structure and stress direction of the overall work area, and then further construct the fault surface. The disadvantages of this manual method are quite obvious: First, it is low efficiency. Due to the work efficiency of the interpreter, the processing time for large-scale work areas is longer. Second, the accuracy is unreliable. The quality of explanation is always affected by artificial deviations. The test results depend heavily on the interpreter's knowledge and experience. Moreover, for areas with less obvious geological characteristics, manual explanation is difficult.
With the development of seismic exploration, geological experts have been tirelessly exploring and researching algorithms to automatically identify seismic faults and helping experts characterize geological structures more accurately.
02 Traditional fault recognition method
Entered the era of computer algorithms. The traditional fault recognition method is based on algorithmic driving of seismic attributes, and uses a single method or model to process seismic data. The main principle is to use seismic properties to calculate faults, integrate information from adjacent seismic channels and samples using a nonlinear way, and directly depict the discontinuity of seismic signals by measuring waveform changes.
1. Early detection methods for faults
The first category is to calculate the continuity attributes of seismic signal reflection to identify faults. The more classic methods include consistency, correlation, similarity, coherence, , emission geometry, eigenvectors, construction tensors, plane wave decomposition and other methods.
The second category is to calculate the discontinuity attribute of seismic signal reflection to identify faults. This type of method has good automatic tracking effect, but is very sensitive to noise and formation characteristics and is prone to false alarms. The main methods include variance , gradient amplitude, chaos, edge detection and other methods.
The third category is to detect the geometric structure of the stratigraphic and use information to interpret the strata to find fault locations. The main methods include stratigraphic inclination, orientation map, etc.
The above method has the following disadvantages. Earthquake attributes usually require a lot of calculations, and local earthquake attributes alone are not suitable for effectively building faults from the overall perspective. They are particularly prone to false positives and require large manual intervention and modification.
2. Enhanced fault recognition method
In the development of traditional earthquake attribute-based algorithms, especially after the emergence of coherent body algorithms, many fault recognition methods with fault enhancement attributes and automatic tracking have emerged one after another, and fault detection is performed through computer assistance.
methods have better effects include: tomographic attribute enhancement, tomography, ant tracking, optimal surface voting, tomography skeletonization, RGB fusion, etc.
fault enhancement methods are prone to two extremes: some methods extract the faults very well, but retain many artifacts ; or the result is clean, but not all faults are detected. There are two limitations for methods like
. First, this approach is difficult to adapt to various discontinuities in different seismic images. For example, coherence techniques are less sensitive to gradually changing faults (i.e., non-sharp faults). Second, the deterministic method itself cannot be systematically learned or developed based on the experiential experience.
03 Modern fault recognition methods
Machine learning (especially deep learning) technologies are very powerful for mining features or relationships from data, which makes them very suitable for learning from human experience, so modern fault detection methods are mainly based on data-driven machine learning.
The first category is machine learning methods. Machine learning is very powerful in data mining and relationship search. It is a data-driven method and is very suitable for learning from human experience.The advantage of the
machine learning algorithm is to find the rules hidden inside the data and be able to efficiently find the mapping relationship between the input data and the target output, which makes the machine learning algorithm very suitable for extracting key information under the conditions of a combination of multiple seismic attributes. Therefore, using machine learning methods for fault recognition can break the limitations of traditional recognition methods and is one of the research hotspots and focuses on intelligent earthquake interpretation.
Classic machine learning methods include: perceptron (MLP) algorithm, Bayesian matching, cross-correlation method, support vector machine (SVM), principal component analysis (PCA), etc.
The second category is the deep learning method.
The main principle of deep learning is a supervised method of extracting meaningful information from a set of inputs (features) and observations (here is fault location), and then applying the acquired knowledge to predict new samples. This algorithm automatically extracts meaningful information from the original amplitude through combinations in the hidden layer of the network, and dynamically generates new features during training. Although training these networks usually takes a lot of time, further output can be obtained very efficiently once the training is completed. There are different ways to apply these methods to fault detection problems in seismic data.
The current deep learning method belongs to the application of artificial intelligence in the geophysical direction and is a popular research field today. Deep networks can dynamically learn new features during training, explaining their success in solving complex tasks.
main methods include BP artificial neural network, CNN convolutional neural network, RNN recurrent neural network, GAN adversarial neural network and other methods. Among them, CNN is the most widely used, including DnCNN residual neural network, UNet neural network and various variants, and the processing effect is relatively good.
Deep learning effect can reach a relatively good level, but there are still some problems with this type of method: First, as a supervised learning method, a large amount of marked seismic data is required as input; Second, generalization problems still exist, and some terrain processing may be good, but the difference will be poor; Third, some special terrain effects are still not good, such as intersection faults, small faults, etc.
04 Future research directions
Future artificial intelligence research is a hot topic, but how to improve processing efficiency, improve recognition accuracy and generalization are all topics that require long-term research. Especially in the processing of three-dimensional data, neural networks occupy a large amount of resources, and the problem of improving processing efficiency is more prominent.
or above is today's course, which mainly explains the traditional and modern technical routes for identifying faults and their respective advantages and disadvantages. I hope it will inspire everyone's research ideas. In the future, some excellent methods will be selected to explain practical cases. That’s all for today’s course, goodbye.
Extended reading:
How to easily get started with seismic exploration research: Start with seismic data processing
How to start the path of earthquake deep learning from 0
