Prediction model for execution time of a production step (timeseries)

jeroenheijlen

Hi,

I'm pretty new in ML and RM and I tried to simplify the situation in the description below.
I want to do predictions towards 'production step execution time' for a low-volume / high-mix environment.

Many different products (P1, P2, ....) are manufactured on production tools (T1, T2, ...) according to their own ‘Product flow’. Some products are manufactured only once, some for years; some products have a production flow existing out of only a few production steps, others have very long production flows.

The production flows are defined by different Productgroups (G1, G2, ...). 

Each production step has its own UniqueStepId and execution time (in minutes). The starting time of each of the production steps is logged (time series).

 A speed-level (slow-normal-fast) is given per step to indicate to the urgency to the operator. Another meta data field is the occurrence of an ‘issue’ during the production step or not (logged by the operator).

 See image

the table shows data for 4 different products (P1-P4) 

of course this is only a small part of the Product flows listed

P1: .... - T4 - T6 - T19 - ...

P2: ... - T6 - T7 - T3 - T9 - ...

P3: ... - T12 - T6 - ...

P4: ... - T1 - ...

 

Questions:

Based on 2019 data,

-        can I create a prediction model for the execution time for a step given the combination ProductGroup-Speedlabel-Tool?

-        can I create a prediction model for the occurrence of an issue yes or no?

Telcontar120

This is all about structuring your data into the correct format.  For instance, with the data structure shown, you could model whether there is an occurrence, at the level of each unique step (not at the product level).  If you want to look at things at the product level, you would need to aggregate your data.  
For execution time, your problem is more complicated.  If you want it at the level you have mentioned, you will need to restructure and aggregate your data so that each row corresponds to the combination of Product-Speed-Tool that you want to analyze.  Based on the example data you have shown, this would be a bit complicated, but not impossible. 
The main thing to consider is that in a normal ML algorithm, each row corresponds to a unique case that you want to examine and has its own independent outcome that you are trying to model.  The rest of the attributes are then the items that are available to the model to make the prediction.  This is what will determine how you aggregate and shape your input data.