Desenvolvimento de um software de visão computacional para estudo do escoamento de cascalhos em peneira vibratória

Abstract

Oil is a fossil fuel of great commercial interest and can be found onshore and offshore. The drilling process has its complexity increasing proportionally to the good depth, being subject to operational problems that can put at risk the integrity of the installations, the operators, and the environment, besides causing non-productive time. A poorly exploited indicator of instabilities is the gravel reaching the surface. Information such as size distribution, speed, and volume with which they return from the well can suggest problems such as crumbling and loss of circulation. The first equipment to receive the drilling mud is the vibrating sieve, and therefore, this study will be directed to this equipment. The Computational Vision is a technology that has as a purpose the automation and integration of processes through the information contained in videos or images using principles of the biological sense of vision. In the last years, several works focused on the oil industry have been dedicated to the applications of these techniques. The objective of this work is to apply computational vision techniques, image processing, and artificial intelligence in the analysis of gravel in vibrating screens, to develop an application capable of monitoring flow variables and detect possible instabilities. Several different materials were used, divided into five groups: i) corn, rice, beans, and peas; ii) artificially colored quartz fragments; iii) quartz fragments, marble, granite, and sand; iv) royal gravel from drilling rigs; and v) 0.1% sand and xanthan gum suspension. Each group used different segmentation techniques for object detection, among them, k-means and Fourier and Wavelets Transform. Estimates of sieve screen-filling and flow velocity were validated, as well as measures such as size, circularity, roundness, and eccentricity of solids. Some matching detectors were studied, and the best performers for this approach were MSER, Harris, and SURF. Texture characteristics were also evaluated as possible qualitative mass flow estimators. Also, convolutional neural networks were evaluated for solids classification when size, composition, and sieve filling. The limitations found were corrected according to the complexity of the study groups. The results show that the application is versatile, and can make good estimates for a variety of materials. In most cases, it was possible to estimate the percentage of sieve filling and the flow velocity with errors of less than 10% considering average values. The study provided great learning about the cited techniques, and the promising results suggest that it is possible to implement a gravel monitoring system in real-time.