Blog
Automated well log correlation: An approach using traditional and FMI logs
'This study was presented at EAGE Workshop on Borehole Geology in Asia Pacific. Borehole images are used by geologists to detect several weak points in wells. Early detection of some patterns,
- legacy-import
- blog
This study was presented at EAGE Workshop on Borehole Geology in Asia Pacific. Borehole images are used by geologists to detect several weak points in wells. Early detection of some patterns, like fractures and breakouts, can prevent a possible collapse of wellbores. Formation micro-imaging (FMI) is one such tool that produces micro-resistivity images of the sidewall of the well bore. FMI logging is used in conjunction with conventional well logging techniques (e.g. GR, Gamma Ray/RES, Resistivity/NPHI, Neutron Porosity/SONIC, Velocity tools) (Watton et. al, 2014). They are also extensively used for structure and texture analysis, fracture evaluation and reservoir characterization.
In order to automate the detection of boundaries/edges in well logs for correlation, established deep learning models in the field of image processing are deployed due to their versatility.
Machine learning is becoming increasingly prevalent for reservoir characteristics analysis in the petroleum industry. Traditional machine learning models require annotated datasets that are interpreted by experts. In reality, researchers often annotate their own training and testing datasets which is a time consuming process (Alaudah, 2019). To overcome scarcity in annotated data researchers have used a) weakly-supervised learning approaches [10], b) similarity based data retrieval [1] and, c) weakly-supervised label mapping algorithm d) Unsupervised image segmentation. We introduce the use of unsupervised image segmentation.
Preprocessing (A) Section of FMI log with white space that represents -9999 and replaced with NaN. (B) NaN removed by removing the columns. (C) Image is converted to gray scale(D) Binary Thresholding of original image (E) Canny edge detection algorithm applied to thresholding image followed by Hough transformation with threshold of 200 and 255 (F) A Classical curve fitting algorithm is used to detect bedding plane.
We use the joint learning approach of Kim et., al , 2020 where the model predicts, for an arbitrary image input, unknown cluster labels and learns the optimal CNN parameters for the image pixel clustering. Subsequently, a group of image pixels in each cluster as a segment is extracted. In the absence of any training images or ground truth labels of pixels specified beforehand, the proposed CNN assigns labels to pixels that denote the cluster to which each pixel belongs thus enabling unsupervised segmentation.
In this paper, we present an approach that exploits a robust unsupervised segmentation of borehole images segmentation to identify bedding planes and fractures.
METHOD
The problem that is solved for image segmentation is described as follows. For simplicity Preprocessing: The detailed overview of preprocessing steps are mentioned below.
- The data contained the values -9999 for the missing data which led to large gaps in the fmi logs. This created problems during the analysis. So, to narrow the gap, we first replaced the values -9999 with NaN and then removed all the columns which had all NaN.
- Grayscale conversion: The image is converted in black and white.
- Thresholding is the simplest method of segmenting images. From a grayscale image, thresholding can be used to create binary images
- The Hough transform is a feature extraction technique used in image analysis, computer vision, and digital image processing. The purpose of the technique is to find imperfect instances of objects within a certain class of shapes by a voting procedure.
- The data contained the values -9999 for the missing data which led to large gaps in the fmi logs. This created problems during the analysis. So, to narrow the gap, we first replaced the values -9999 with NaN and then removed all the columns which had all NaN.
The Network Architecture : Training CNN to extract features then, calculating the response vectors and normalizing it. Argmax function assigns the cluster labels to the pixel
Input image is fed into the CNN to extract deep features using a feature extraction module. Subsequently, one-dimensional (1D) convolutional layer calculates the response vectors of the features in q-dimensional cluster space, where q = 3 in this study. Subsequently, the response vectors are normalized across the axes of the cluster space using a batch normalization function. Further, cluster labels are determined by assigning the cluster IDs to the response vectors using an argmax function. The cluster labels are then used as pseudo targets to compute the feature similarity loss. Finally, the spatial continuity loss as well as the feature similarity loss are computed and backpropagated
(A) Original image Training applied for 500 iterations without hyper-parameter training which detects 20 clusters (C) Grayscale version of the original image. (D) 19 clusters identified by gray scale images.
(A) Original image (B) Training applied for 500 iterations with µ=0.1 which detects 31 clusters (C) Training applied for 500 iterations with µ=0.5 which detects 31 clusters (D) Training applied for 500 iterations with µ=1 which detects 20 clusters. (E)Training applied for 500 iterations with µ=2 which detects 12 clusters (F)Training applied for 500 iterations with µ=3 which detects 19 clusters
(A) Original image (B) Training applied for 50 iterations with µ=10 training which detects 11 clusters (C) Handmade Scribble of the original image (D) Gray scale version of the original image (E) 11 clusters identified after using scribble.
In conclusion, CNN based unsupervised learning is able to create clusters and identify sinusoidal patterns of fractures and bedding planes. This method needs to be verified for complex fractured regions .
