A physics-constrained unsupervised machine-learning method for collar wave suppression in acoustic logging while drilling

Acoustic logging while drilling (ALWD) plays a pivotal role in determining the P-and S-wave velocities of formations around the borehole during drilling operations. However, the extraction of formation P-wave velocities in monopole measurements faces significant challenges due to strong interference from collar waves, which travel along the tool body. Common industry solutions involve using an isolator to suppress these collar waves, but this approach can compromise the stiffness of the collar and lead to engineering challenges. This study introduces a novel approach to suppress collar waves using a physics-constrained unsupervised machine-learning method. This is the first time that move-out differences between collar waves and formation P waves have been used to suppress collar waves in ALWD effectively. We evaluate the effectiveness of this method on synthetic data and a scaled laboratory experiment by measuring the P-wave velocity errors, demonstrating its superior performance over traditional median filtering. In the synthetic tests, we assess formation P-wave velocities ranging from 3200 m/s to 5000 m/s, in increments of 200 m/s, under two different outer fluid radii conditions (100 mm and 110 mm). The results show that the average errors between the actual and estimated P-wave velocities, after applying the unsupervised method for collar wave suppression, are 1.0% for an outer fluid radius of 100 mm and 1.1% for 110 mm. In contrast, median filtering results in average velocity errors of 30.5% for a 100 mm radius and 30.8% for a 110 mm radius. In a scaled laboratory experiment, the P-wave velocity extracted using the unsupervised method, after collar wave suppression, is 3860 m/s, resulting in a minimal velocity error of 1.0%. However, the P-wave velocity extracted using median filtering is only 2400 m/s, leading to a substantial error of 38.5%.