DeepSeg-based noise reduction algorithm trained on a hybrid synthetic dataset for signals from acoustic logging-while-drilling

Acoustic logging-while-drilling (ALWD) enables real-time acoustic measurements during drilling operations. However, challenging downhole conditions introduce considerable noise into ALWD signals. This study applied a numerical method to simulate clean ALWD array waveforms. A synthetic noisy dataset was subsequently generated by superimposing high-quality noise data acquired during laboratory measurements onto the simulated clean dataset. Time–frequency spectral representations of noisy signals were obtained via short-time discrete cosine transformation and were divided into past, present, and future intervals. These were used as input for a U-Net-based neural network—DeepSeg—that was employed for frequency-domain denoising. The trained model outputted denoised frequency segments for the intermediate current time interval. This sliding-window strategy applied to frequency slices substantially reduced the required dataset size. The network effectively removed complex downhole noise, even with limited training data, and demonstrated a strong generalization capability on field data. The network also significantly enhanced the quality of acoustic array signals and improved the accuracy of slowness-time-coherence processing in a logging-while-drilling application example. Requiring as little as one tenth of the amount of data used in methods from previous studies, the proposed method is particularly advantageous for ALWD applications, where data acquisition is challenging. • A numerical method for simulating array ALWD waveforms is introduced. • Frequency-domain denoising was performed using a U-Net-based neural network. • Complex downhole noise is effectively denoised even for limited training data. • The quality of array ALWD signals is significantly enhanced. • The accuracy of Slowness–Time–Coherence processing is improved.