A modular benchmark simulator for Flexible Assembly Systems (FAS) or Flexible Manufacturing Systems (FMS).

Licence: WTFPL

This project includes a SimPy discrete event simulator simulating a plausible Flexible Assembly System documented in the documentation/FAS-Simulator.pdf.

See Executive summary poster in: documentation/ExecutiveSummary.pdf.

It contains several pre-configured scripts for different kinds of production runs, named: run*.py

See Python dependencies from requirements.txt. For plotting things you need Octave and epstool.

Example: Running a simple simulation with a simulated wear and tear fault: ./run_easy_with_wear_and_tear_fault.py

Running the simulations produces output to the STDOUT, but the actual output is written as JSON to output.json.

The output contains a sequence of events with timestamps.

Additionally, there are several *.m files and *.sh files in utils directory to create different kinds of visualizations out of this JSON output using Octave and ffmepg.

The data.mat file for the Octave scripts is created using ./utils/output_to_octave.py from output.json to data.mat.

There is a baseline implementation using Hidden Markov Models for anomaly detection in: fas_hmm

There is also a related FAS-Tensorflow project (private at the moment) that is an implementation that extracts process model features from inputs generated by this project using deep learning methods, and uses those for anomaly detection. This is about a novel deep learning architecture which leverages interlaced sequential symmetry in the data.

There are several runs implemented:

• Run without errors
• Wear and tear fault, where a Conveyor is affected by a progressive delay.
• Retry delay fault, where a BowlFeeder process suffers from failed feed randomly, but continues correctly when retried.

For the challenge task we create a set of clean runs, and a set of runs with errors happening at the starting point in the run.

A suggested learning system can be trained or conditioned on a subset of clean runs, and tested whether it is able to discriminate between runs with errors and clean runs. Note that a system shouldn't be trained on the faulty runs, because the fault types and behaviors cannot be known in advance in practice.

We first produce 10,000 runs with errors and 10,000 similar runs without errors in JSON.

Do this by running: ./generate_datasets.py.

The data will be generated in JSON form, one file per run in data/correct_runs/*.json and data/runs_with_errors/*.json.

This data can then be converted to numpy format by running: ./convert_datasets_to_numpy.py.

The data will be generated in numpy form, one file per run in numpy_data/correct_runs/*.npy and numpy_data/runs_with_errors/*.npy.

FAS Simulator

@article{keski2017simulator,
  title={A simulator for event-oriented data in flexible assembly system fault prediction},
  author={Keski-Valkama, Tero},
  journal={Procedia computer science},
  volume={119},
  pages={121--130},
  year={2017},
  publisher={Elsevier}
}

• https://github.com/keskival/FAS-Simulator