Machine Learning for Source Location

More details to follow. For the moment there are just two component codes.

blockdata.py

A digital twin for a lab rig.

To solve a 3D linear quasistatic or viscodynamic problem using python and FEniCS http://fenicsproject.org. The right hand side forcing is narrowly located at a randomly chosen point in the domain, vibrating at a carrier frequency which is then modulated by a lower frequency. The response is then picked up by microphones and accelerometers at given points on the boundary.

The bar has dimensions 30cm long, by 10 wide and 5 deep. The bottom is considered to be fixed in position for all time (homogeneous essential - Dirichlet - boundary condition) and the remaining faces are traction free (homogeneous natural - Neumann - boundary condition).

Based on demo_elastodynamics.py as downloaded on 10 Dec 2019 from https://fenicsproject.org/docs/dolfin/latest/python/demos/elastodynamics and also on ft06_elasticity.py as downloaded on 10 Dec 2019 from https://github.com/hplgit/fenics-tutorial/blob/master/pub/python/vol1/ft06_elasticity.py

The code is awaiting many mods:

• addition of a quasistatic viscoelastic capability using internal variables
• addition of a quasistatic viscoelastic capability using fractional calculus 'power laws'
• addition of at least one form of time stepper for the dynamic problem(s)

Eventually this will be used to generate virtual training data for a machine learning source identification problem, as well as the test and validation data. We would like more than one forward solver in order to address so-called inverse crime

postpro.py

To post process sensor data from blockdata.py and compress it to a form suited to training stage the Neural Net.

The code is awaiting many mods:

• intelligent and variable interfacing to the blockdata.py code so that post-processing is seamless.

ihcc_linear_elastostatic_exact_data.py

Contains FEniCS expressions for generating manufactured exact solutions for checking consistency/convergence.