Fuzzy Neural Network for Studying Coupling between Drilling Parameters

Abstract

The rate of penetration (ROP) is an index used to measure drilling efficiency. However, it is restricted by many factors, and there is a coupling relationship among them. In this study, the random forest algorithm is used to sort influencing factors in order of feature importance. In this way, less influential factors can be removed. A fuzzy neural network (FNN) is applied to the field of drilling engineering for the first time, aiming at the coupling problem to predict the ROP. Fuzzification is an important part of training and realizing FNN, but research on this topic is currently lacking. In this study, K-means are used to divide the data with high similarity into a fuzzy set, which is used as the initialization parameter for the second layer of the FNN. The data of Shunbei No. 1 and 5 fault zones in Xinjiang are collected and trained. The results show that the mean value of the coefficient of determination R 2 is 0.9668 under 10 experiments, which is higher than those obtained from a back propagation neural network and multilayer perceptron particle swarm optimization methods. Therefore, the effectiveness and feasibility of the model are verified. The proposed model can improve drilling efficiency and save drilling costs.

Improving the Drilling Parameter Optimization Method Based on the Fireworks Algorithm

“…Previously, the ROP equation and mechanical specific energy − have been the main methods to establish the optimization index (i.e., the fitness function), but because of their excessive dependence on formation information, they require engineers to give the corresponding parameter settings through experiments, which brings considerable difficulty to optimize the ROP. With the popularization of big data analysis methods, ROP prediction models based on machine learning and deep learning have gradually become new research techniques, which calculate and generalize the laws existing in the data to reach the predicted ROP values under similar conditions by simply feeding the controllable parameters into the trained model . For example, Ahmed et al used an artificial neural network (ANN) to predict ROP.…”

“…With the popularization of big data analysis methods, ROP prediction models based on machine learning and deep learning have gradually become new research techniques, which calculate and generalize the laws existing in the data to reach the predicted ROP values under similar conditions by simply feeding the controllable parameters into the trained model. 18 For example, Ahmed et al 19 used an artificial neural network (ANN) to predict ROP. Kor et al 20 compared different prediction methods based on a statistics viewpoint and proved the effectiveness of support vector machine regression in ROP prediction.…”