A deep learning model for subsurface stratigraphic classification with continuous CPT data

Abstract

The cone penetration test (CPT) records cone tip resistance and sleeve friction continuously. Relying on this data feature, this paper proposes a refined bidirectional-head-cohesion long short-term memory (BHC-LSTM) model for accurate data stratigraphic delineation by integrating the new concept of overlapping frames and two LSTM networks. The novel BHC-LSTM method uses the information from both above and below the target soil layer simultaneously for refined soil classification to improve prediction accuracy. The model performance is examined using self-measured data from two engineering sites in Jinan and a published database, achieving an overall prediction accuracy higher than 95%. The results show that the BHC-LSTM model can significantly improve the prediction efficiency and accuracy for stratigraphic soil types compared with other conventional deep learning methods. The new method can benefit soil layer classification based on CPT data in geological surveys.

Affiliated Institutions

• Shandong Transportation Research Institute CN
• Qilu University of Technology CN
• Shandong University CN
• Shandong Academy of Sciences CN

Related Publications