#hackingforfuture

The Atlas Copco Group serves customers through innovative compressors, vacuum solutions, generators, pumps, power tools and assembly systems. We are a global and diverse Group of many strong brands and around 34 000 employees representing different cultures in more than 180 countries.

The Airtec division is the heart of the compressor technique business area, because in Airtec we manufacture and assemble the core elements of our compressors. In our factory we have operators working in 3 shifts, to keep the machines running 24/7. The installed base includes turning lathes, 5 axis workcentres, millturns, grinding machines and robots, mainly machining steel and cast iron parts. This makes the inside of the factory a real industrial environment, something to keep in mind during the hackathon.

For this hackathon we want to create a digital operator assistant, focusing on triggering and encouraging the operator to provide feedback on why certain machines are not running or have stopped. This feedback should then be stored together with the source event in a database for later analysis.

This will include following tasks:

Pick the platform

Origins: connecting to the source

End game: connecting to a sink source and back end

Big Brother option: connector for third party real time acces

The holy grail: create the best possible GIU for operators

The not so holy grail: create admin front end and supervisor dashboard.

The sink-EDM process is not that describable by machine data due to its specific characterization: The electrode is removing the material based on electronic discharge and you never know where material is removed. This problem can be solved with data! You are given a set of machine data that resulted from a sink-EDM process. Your goal is to allocate the quality relevant numerical values on the surfaces of the virtual CAD model of the work piece. The results should be presented in a work piece »Digital Twin«: The data points have to be visualized and help the operator to understand the process much better and thus increase its quality.

This will include the following tasks:

The first step is to do some good old coordinate transformation to bring the measurements into the work piece position

Now do some head over heels and map the numerical data from the electrode surface to the work piece surface

For a better analysis of the data at a later stage, you have to decide in which data base format you want to store the data and how to ensure to store the results in an AAS data structure

Finally be the artist and visualize the data on the demonstrator geometry

How nice would it be to have a central intelligence in production that is monitoring and analyzing arbitrary process parameters automatically with regard to given specification limits of these process parameters?

This is exactly where the trend is going to and you can contribute to that together with us!

Therefore we will provide you with labelled data sets showing values of process parameters over time. The according plots and the histograms of these process data sets show certain characteristics or patterns that you shall be able to classify.

This will include the following tasks:

Classify if the location and/or the scattering of a process is changing over time (labelled data)

Count the number of peaks in a histogram

Classify the kind of probability distribution based on the histogram (normal distribution or something else?) (labelled data)

Count the number of different probability distributions mixed together and classify them

Identify jumps, outliers and trends along a curve of process parameter values over time

What if you could predict the quality of the parts in production already during production?
Then, you would only have to measure the parts that are predicted to be critical or you can improve your processes, or you could just skip the quality checks at the end. This would save resources, time and money.

Try to get out the most of the data we will provide to you and predict the quality of the underlying parts!

The structured data consists of columns representing input parameters and columns representing output parameters. Try to predict the output parameters based on the input parameters by regression and/or classification analyses!

This will include the following tasks:

Pre-process the given data (different files representing different processes with different inputs and outputs)

Implement different approaches (e.g. decision trees, gradient boosted trees, neural networks) to predict the outputs

Study the accuracy and the computational performance of the best approach for each given data set

Still not enough? Can you make use of automated machine learning libraries (e.g. Auto-Kera, Auto-Sklearn) to find the best approach for each data set automatically?