Accurate prediction of attitude angles in the measurement while drilling system through backpropagation neural networks

Abstract Directional drilling machines are extensively employed in gas drainage, water exploration, and coalbed methane exploration in underground coal mines. The measurement while drilling (MWD) system collects the spatial orientation and depth parameters of the drill rod during the drilling process using a three-axis sensor, which guides the drill rod along the predefined trajectory, ensuring it reaches the designated drilling position for safe coal mining. Due to installation angle errors of the three-axis sensor and alignment discrepancies between sensors, the sensor produces complex three-dimensional spatial errors during measurement, complicating the identification of the physical plane correction axis needed for traditional installation error correction methods. To resolve this issue, a BP neural network-based correction algorithm for installation errors in the MWD system is proposed. The algorithm disregards installation errors of accelerometer and magnetometer sensors, does not necessitate additional parameter calculations, and directly predicts attitude angles through ellipsoid fitting of the measurement data, thereby effectively mitigating the impact of installation errors. The effectiveness and prediction accuracy of the algorithm are validated through numerical model experiments, ground experiments, and underground tests. The results demonstrate that the absolute azimuth error in the indoor experiment is less than 1°, improving by approximately 5° compared to the traditional scalar product constant method, with an absolute pitch angle error of 0.1°, improving by approximately 1.5° compared to the traditional plane-fitting method. The absolute azimuth error in the ground experiment is less than ±1.8°, and the absolute pitch angle error is less than ±0.18°, which meets the requirements of the while-drilling measurement system for underground coal mining projects.