- Problem
- Laboratory hydraulic fracturing experiments produce hours of multichannel acoustic emission recordings, while each event lasts a fraction of a millisecond. Events must be detected, their arrivals picked and their sources located before the data says anything about how the rock cracked.
- Approach
- An interactive workflow: load the recordings, let a picker find the arrivals, correct picks directly on the trace, invert the picks for a source position, and set the result against the injection pump record.
- My contribution
- Built the application, including a PyTorch convolutional neural network for acoustic emission detection and picking, and the review, localization and pump-comparison workflow.
Overview
HydraFrac Studio is the interactive analysis tool for the acoustic emission (AE) data recorded in the HydraFrac laboratory hydraulic fracturing experiments. It takes a researcher from raw recordings to picked arrivals, located sources and a comparison with the injection record. It runs locally, so research data never leaves the researcher's machine.
Problem addressed
An AE event in a rock specimen lasts a fraction of a millisecond, but a continuous recording runs for hours across many channels. Before any interpretation is possible, the recording has to be cut into events, the first arrival on each channel has to be picked precisely, and the picks have to be inverted for a source position. Doing this by hand does not scale, and fully automatic picking cannot be trusted without review.
Technical approach
- Picking. A deep learning picker is the default, with classical, interpretable pickers alongside it. The classical methods need no training and act as an independent check on the network.
- Review. Researchers can correct any pick directly on the trace, and their corrections are trusted in the later steps.
- Localization. Picks are inverted for source positions and refined with waveform-based imaging.
- AE against the pump record. The AE timeline is aligned with the injection record, to show when the specimen started to crack, at what pressure and after how much fluid.
- Testing. Automated tests check the workflow end to end against synthetic experiments with known answers.
My contributions
I built the application: the deep learning picker, the interactive review interface and the localization and pump-comparison workflow.
Results
The tool is used internally by researchers in the HydraFrac group. On real triggered records, the deep learning picker and the AIC picker, two independent methods, agree to within 2 µs on 69% of the picks both are confident about, and to within 10 µs on 92%.