NTNU (Trondheim, Norway) collaborates with many industries: telecom, construction, shipyards, engineering, which strive to incorporate machine learning into their business and operation processes. One of such companies is TechnipFMC that develops surface, subsea, off- and onshore solutions for the energy sector. Its cooperation with NTNU includes research and co-supervision of master and PhD students.
For TechnipFMC the problem of estimating the fatigue of drilling risers, which connect the well with the surface, is crucial. They are consumables and should be replaced in time to prevent dangerous failures and halts of operation. On the other hand, replacing risers too early leads to financial losses. Manufacturers and operators install sensors (usually, accelerometers) on risers to monitor their state.
This is when machine learning comes onto the scene. Having these measurements, companies want to predict the fatigue of the wellhead a riser is connected to (with an appropriate model model it is also possible to estimate the wear of the riser itself). The sensors, however, also require maintenance, and the question is to place them in such a way that it is easy to fix them (remember the mining happens in a cold sea) while the readings are representative.
In our work, we wanted answer several questions: How useful can machine learning be in predicting the fatigue of slender structures (like marine risers and beams)? How does the complexity of a structure’s dynamics affect prediction results? Where should a few sensors be placed on a structure? Our hypothesis was that for structures like a simple Euler–Bernoulli beam linear ML models should produce good predictions, while the sophisticated dynamics of a drilling riser would require complex models. A secondary objective was to try (and verify in some sense) novel neural network architectures inspired by differential equations.
Our experiments showed that linear regression is enough to predict the maximal bending moment of a clamped Euler–Bernoulli beam and that sensor positions do not affect the results significantly. However, for a drilling riser —even described by a simplified set of equations— a so-called Antisymmetric Recurrent Neural Network offers notably better prediction when only two conveniently placed sensors are used.
We can draw two conclusions from the results. Firstly, given the amount of data that offshore engineering companies already have and can obtain in the future, ML is a promising approach that can complement traditional methods. Secondly, ML models for mission-critical applications should be developed further to guarantee robust prediction, and the mentioned approach based on differential equations is one of the most promising directions.