Formation Waves Decoupling Network: Enhancing Cementing Quality Evaluation in Fast Formations

Evaluating cementing quality is crucial in the construc- tion of oil and gas wells. However, this process is partic- ularly challenging in formations with high acoustic wave speeds (fast formations) due to interference from forma- tion signals and the scarcity of specific formation data while logging. The rapid propagation speed of acoustic signals in fast formations causes overlap between forma- tion waves (FW) and other waves, mainly casing waves, in the Variable Density Log (VDL). This overlap makes it difficult to distinguish between the two signals manu- ally, leading to low evaluation efficiency. Furthermore, the limited amount of data available for fast formations makes it challenging to use supervised machine learning methods for evaluation. To address these challenges, we propose a Formation Waves Decoupling Network (FWDNet) designed to sep- arate FW from VDL in fast formations. The model com- prises a VDL encoder, a FW decoder, a residual waves (RW) decoder, and a signal mask module. We utilize open-hole logging data, including the Acoustic Log (AC) and Gamma Ray (GR), to introduce original formation information. The open-hole and cased-hole logging data are input into the signal mask module to obtain the mask vector of FW, which is then multiplied by the VDL em- bedding vector to decouple the formation signal from VDL in the embedding space. Subsequently, the FW and RW are reconstructed by decoders. In the first training phase, we perform frequency-domain signal separation on the Unmixed VDL of conventional formations and revert them to time-domain signals. This process con- structs labels for the FW and RW for supervised train- ing. In the second training phase, we fix the weights of the encoder and decoders and use the mixed VDL from fast formations to obtain the FW and RW which are com- bined to reconstruct the VDL. Self-supervised training is used to optimize the signal mask module for fast forma- tion. Finally, experiments on real logging data demon- strate that using decoupled FW and RW as training data significantly improves the accuracy of cementing quality evaluation in fast formations compared to models based on original VDL. This showcases the effectiveness of the proposed method for decoupling formation signals from VDL in fast formations.