Collaborative multi-agent rock facies classification from wireline well log data

Evaluation of automated lithology classification architectures using highly-sampled wireline logs for coal exploration

“…Prominent among the methods used with wireline logs are Fuzzy logic (Hsieh et al, 2005), Naïve Bayes (NB) classifiers (Li and Anderson-Sprecher, 2006), Artificial Neural Networks (ANNs) (Qi and Carr, 2006;Dubois et al, 2007;Al-Anazi and Gates, 2010;Wang and Carr, 2012a), and Support Vector Machines (SVMs) (Al-Anazi and Gates, 2010; Wang and Carr, 2012b). Although commonly used independently, the use of heterogeneous committees of classifiers has also been investigated (Gifford and Agah, 2010). The accuracy of these algorithms have in many cases been compared to that obtained with Linear Discriminant Analysis (LDA).…”

“…These applications utilised wireline logs which were sampled at 15 cm (Qi and Carr, 2006;Dubois et al, 2007;Gifford and Agah, 2010); 30 cm (Chang et al, 2000(Chang et al, , 2002; 50 cm (AlAnazi and Gates, 2010); and 2 m (Hsieh et al, 2005). The number of wireline logs used (i.e.…”

“…the data's dimensionality) also varied, ranging from 4 to 9. In many cases the classifiers were evaluated on held-out depth intervals from the training well (Qi and Carr, 2006;Dubois et al, 2007;Gifford and Agah, 2010;Hsieh et al, 2005), which may not give performance that is representative of use on unseen wells (Dubois et al, 2007).…”

Extreme Learning Machine combined with a Differential Evolution algorithm for lithology identification

“…Machine learning approaches can potentially make the process of reservoir and rock formation identification more efficiently by providing the means to formalize the expert knowledge through know-how engineering [7]. Some research efforts found in the literature are described as follows.…”

“…Horrocks et al [2] explores different machine learning algorithms and architectures for classifying lithologies using wireline data for coal exploration. Other approaches include multivariate statistical analysis [11], neural networks with probabilistic neurons [12] or radial basis function kernel [13], random forests [14,15], combination of classification and regression methods [16] and collaborative learning agents [7]. ELM networks may need a higher number of hidden neurons due to the random determination of the input weights and hidden biases.…”