This invention is an artificial intelligence method that trains a neural network on a combination of simulated and experimental guided-wave data to predict whether a location along a pipeline has a defect, such as corrosion. It uses adversarial training so that models built largely on simulation data can work on real pipeline measurements, which reduces the need for large sets of labeled experimental data.
Description
The technology combines computer simulation of guided-wave (ultrasonic) propagation along a pipeline with real measurements taken from the pipe, including measurements from multiple fiber sensors. Simulations use finite element analysis together with reduced-order modeling, a compression technique that generates large numbers of simulated signals quickly. A convolutional neural network can then combine the simulated and experimental signals into synthetic data, represented as time-frequency images called continuous wavelet transform (CWT) spectrograms.
A neural network is then trained on this data using an adversarial training technique. In adversarial training, the model learns features that are indistinguishable between simulated and real data, so it can transfer from simulation to the real pipeline. The trained model predicts whether a pipeline location has a defect and can classify the location as a weld or as corrosion. The trained model is then used to run inference tasks on pipeline data.
Applications
- Oil and gas pipeline integrity monitoring and inspection
- Permanently installed guided-wave and fiber optic structural health monitoring systems
- Corrosion detection for utility and energy infrastructure
- Non-destructive evaluation software and services for infrastructure operators
- Inspection of other large structures, such as bridges and nuclear fuel systems
Advantages
- Trains on simulated data alone or combined with experimental data, reducing dependence on large labeled experimental datasets
- Uses adversarial training and domain adaptation to narrow the gap between simulation and real measurements
- Distinguishes weld locations from corrosion locations, which helps limit false corrosion calls
- Reduced-order modeling combined with finite element analysis lets large simulated datasets be generated quickly
- Works with data from multiple fiber sensors, supporting distributed monitoring
Invention Readiness
The technology has been implemented and tested as a working prototype framework using simulated pipeline data and experimental data collected from a physical carbon steel pipe instrumented with fiber optic sensors. It detects excitation, weld, corrosion/defect and pipe-end reflections.
IP Status
Patent Pending