SLIDE

The method SLiding-window Initially-truncated Dynamic-response Estimator (SLIDE) is a deep-learning based method for estimating the output of mechanical and multibody systems.
The corresponding research paper is now available on arXiv, doi: arXiv.2409.18272, and in the submission process for journal publication.

In the development following versions of packages were used:

• Python 3.11.8
• pytorch 2.2.1 (+cu121)
• exudyn 1.8.52 (and newer)
• matplotlib
• numpy 1.23.5
• ngsolve 6.2.2403

The according requirements.txt file can be used to install the requirements with
pip install -r requirements.txt
For the flexible 6R robot example additionally ngsolve is required. We used Version 6.2.2403.

Exudyn

Exudyn is a flexible multibody dynamics simulation code. The C++ core ensures efficient simulation, while the Python interface enables compatibility with pytorch and other machine learning tools. For more information see the extensive documentation and the examples on github.

Source files:

Please note that the provided source code is experimental research. There are both models in the simModels and features in the AISurrogateLib which are not part of the publication and may not work. The object NeuralNetworkTrainingCenter is a container object for both simulation and neural networks. It is denoted nntc in the code and enables parameter variations, thus training a number of neural networks in parallel with different parameter settings.

• simModels: here the class definitions for the simulation models and some helper / utility functions are implemented.

  • SimulationModel: parent class from which we derive all other simulation models.
  • Oscillator: both linear and Duffing oscillator, depending on the provided parameters.
  • OscillatorContinuous: same as oscillator, but truncation is done directly to the data.
  • SliderCrank: Slider-crank system. By setting the flexible flag the connecting rod is modeled with ANCF elements. Also rigid body simulation can be switched on, but then the output has to be changed to outputType=0, which measures the slider's position.
  • Flex6RRobot: The puma 560 standing on a flexible socket.

• AISurrogateLib: The base library providing "presets" for neural networks. In the experiments we can derive custom neural networks as e.g. for the flexible robot example.

Examples

Please note that all examples also create the associated data from the simulation. While the oscillators run very fast, creating data for the flexible system may take longer. In the paper, data creation was done partly on a cluster (Leo5), while the training itself was performed locally. Data is only created if the according file does not exist yet.

• 01_LinearOscillator.py: learning of linear oscillator system.
• 02_DuffingOscillator.py: learning of Duffing oscillator including the error estimator using the SLIDE method.
• 03_SliderCrankFlexible.py: the flexible slider-crank system including error estimator.
• 04_Flexible_6RRobot.py: the Flexible Robot including error estimator.

Source/data:

In this directory, the training data for the supervised learning of the neural networks is saved.

Source/model:

In this directory, the parameters from the neural networks are saved as *.pth files and can be loaded either with the utility function of the neural network training center LoadNNModel.

Source/solution:

Here training and validation loss (MSE) is saved for plotting.

Licence

See Licence.txt.