Investigating the Effect of Different Noise Reduction Filters on the Rate of Penetration Approximation by Machine Learning Based Models

Accurate modeling and estimation of drilling rate of penetration (ROP) is a prelude to better planning to reduce drilling time and control costs. However, estimating this key factor is not easy and the main reason is the complex relationship between penetration rate and drilling variables. On the other hand, the presence of noise in the data increases the training time of the models and drastically reduces their accuracy. In this study, the multilayer perceptron neural network (MLP) method and the least squares support vector regression (LSSVR) method along with four data noise reduction filters have been used to estimate the penetration rate. The data used in order to feed the models, collected from mud logging unit (MLU) and final report of a drilled well that located in south-west Iran. After the feature selection process with simulated annealing (SA) optimization algorithm and artificial neural networks, eleven variables were selected from all of the variables. Out-of-range points were then removed and the overall noise of the data was reduced by median filter, savitzky-golay, moving average and signal wavelet. Finally, the performance of the developed models and filters used were evaluated and compared by different statistical indicators and it was proved that the LSSVR method with radial basis kernel and output data from moving average filter shows the best results in the both training and testing section of the model