A Conventional Neural Network Lithology Classification Method Based on Vibration Data

Abstract

ABSTRACT: Drilling string vibration data is a high-density ancillary data and it has the advantages of low-latency and low-cost which can be acquired in real time. In this study, vibration dataset is used as signal source, and the original vibration signal is filtered by Butterworth (BHPF). vibration time-frequency characteristics are extracted into time frequency images with the application of short-time Fourier transform (STFT). This paper develops lithology classification models using new data sources based on convolutional neural network (CNN) combining with Mobilenet and ResNet. This model is used for complex formation lithology including fine gravel sandstone, fine sandstone and mudstone. In order to improve the trustworthiness of decision-making results, the gradient-weighted class-activated thermal localization map is applied to interpret the results of the model. The final vertification test shows that the single-sample decision time of the model is 10ms, the test macro precision rate is 90.0%, and the macro recall rate is 89.3%. The lithology classification model is more efficiency and accessible. In conclusion, The CNN model using drill string vibration supplies a superior method of lithology classification. This study provides low-latency and low-cost lithology judgment methods to ensure safe and rapid drilling. 1. INTRODUCTION Lithology classification of underground formation is of great importance in the field of oil & gas exploration engineering as lithology represents the reservoir petrophysical characterization.(Buryakovsky et al. 2012). Vibration data of drill string can also be used to make real time lithology classification considering different formation characteristics. (Esmaeili et al. 2012b) Some researchers made logging-based lithology prediction by using artificial neural networks. It is demonstrated that lithofacies information from images could be used for lithology classification based on ANN.(Ivchenko et al. 2018) Random Forest is always applied to make underground formation lithology compared with other machine learning algorithms such as GTB, GBM and AdaBoost.(J. Sun et al. 2019) Baraboshkin applied convolutional neural networks on rock description based on color distribution and feature extraction by using different neural network architectures. GoogleNet make a better performance than other algorithms.(Baraboshkin et al. 2020) Ahemd used three machined learning models to predict the lithology changes and formation tops in real-time while drilling including ANN, ANNFIS and FNN, the results shows ANN model shows better performance.