Abstrak
Monitoring proses hydraulic fracturing merupakan langkah penting untuk memastikan keberhasilan dan efisiensi stimulasi reservoar. Dalam pelaksanaannya, monitoring dilakukan menggunakan hasil rekaman seismik yang mendeteksi rekahan yang terjadi di ujung bawah wellbore. Sistem ini dibangun untuk mengidentifikasi event mikroseismik akibat rekahan, berdasarkan hasil sinyal seismik yang direkam selama proses berlangsung. Penelitian ini bertujuan mengembangkan sistem deteksi event mikroseismik secara otomatis menggunakan pendekatan deep learning. Konsep transfer learning dimanfaatkan karena telah banyak diterapkan dalam konteks seismologi, seperti pada model pre-trained PhaseNet, U-GPD, dan EQTransformer untuk keperluan klasifikasi event maupun phase picking gelombang P dan S. Penelitian ini bertujuan untuk mengembangkan pre-trained model, khusus untuk data mikroseismik dari wilayah Formasi Talang Akar. Arsitektur CNN yang digunakan, dilatih menggunakan data hasil pengukuran seismometer dari Lapangan X dan menghasilkan akurasi sebesar 75,63%. Lapisan fitur dari pre-trained model ini kemudian dibekukan (frozen) untuk mengekstraksi fitur dari data target, yaitu data selama proses Step Rate Test (SRT) di lokasi yang sama. Fitur tersebut selanjutnya diklasifikasikan menggunakan algoritma machine learning konvensional, dengan pendekatan ini dihasilkan akurasi terbaik sebesar 76,92% dalam mendeteksi event mikroseismik akibat proses SRT.
......Monitoring the hydraulic fracturing process is a crucial step to ensure the success and efficiency of reservoir stimulation. In practice, monitoring is carried out using seismic recordings that detect fractures occurring at the bottom tip of the wellbore. This system is designed to identify microseismic events caused by fracturing, based on the seismic signals recorded throughout the process. This study aims to develop an automatic microseismic event detection system using a deep learning approach. Transfer learning is utilized, as it has been widely applied in seismology, for example in pre-trained models such as PhaseNet, U-GPD, and EQTransformer for event classification and P- and S-phase picking. This research focuses on developing a pre-trained model specifically for microseismic data from the Talang Akar Formation. The proposed CNN architecture was trained using seismic data recorded at Field X and achieved an accuracy of 75.63%. The feature layers of this pre-trained model were then frozen to extract features from the target data, which were obtained during the Step Rate Test (SRT) process at the same location. These features were subsequently classified using conventional machine learning algorithms, resulting in the best accuracy of 76.92% in detecting microseismic events induced by the SRT process.
