| dc.description.abstract | Drilling fluid rheology is important for drilling safety and has to be constantly monitored
and adjusted during drilling operations. The goal of this thesis is to attempt to create a
model with neural networks that can estimate rheological properties of drilling fluids based
on dampening and travel time of ultrasonic waves. The current way of measuring rheology
consists of sampling and use of offline rheological measurements using lab equipment, and
an online measurement system would allow for faster corrections.
Experiments have been planned and carried out accordingly to create data for training and
testing the neural network models, and this data has been used for training the models along
with previously gathered data.
The neural network models have been created with TensorFlow in Python, with Adam
Optimiser, relu6 and sigmoid activation functions, and square error loss function. Models
have been created for Density, Yield Point, Gel Strength and Plastic Viscosity. The best
models for each output, in the same order, have an RMSE of 2.7%, 2.2%, 1.7% and 3.0%
with all available data based on two different drilling fluids gradually diluted, and 5.1%,
3.6%, 3.7% and 3.8% with data gathered in this thesis based on one type of drilling fluid
gradually diluted, where the best models were selected based on mean square error. These
models were the best out of more than 250 models each that were trained with the same
datasets for the same output variable. | nb_NO |
