This week I have been spending time analyzing the amount of Touch Time and the Cycle Time for the production of engineered circuit boards. I thought of something overnight that I had not taken into consideration the first time I had analyzed the "actual" data. There are different stages of the design (for example, new or a relayout) and the type of board that is being created (for example, digital or analog).
This morning, I reran my randomly selected actuals and record the outcomes. I did find that my results did change, and slightly for the better as far as the tool was a little better at predicting with the design stage and type of board information is entered. It was not, however, significantly better and still rather poor at predicting.
I put together the data into a formal statistical test. Using the method we used in my AP Stats class, I drew the conclusion that the tool was at best 15% accurate in predicting the Touch Time and 18% accurate in predicting the Cycle Time. This is not good and is not what my department wants.
I was able to present this information to some of the people this tool effects. Their instincts had told them the tool was not good, but this was verification of that instinct.
I then showed them why the tool is not good. The reason is not the tool. The reason for the inaccuracy is the inconsistency of the actual data. I showed them several of specific instances of these inconsistencies. So I drew the conclusion that the current information entries into the tool do not correlate well with the two measures of time. I found some better correlated data, but I'm not convinced even having those data predicted would give a better result.
Overall, I think they liked my presentation of the data. There were a few additional ideas to come out of the meeting: 1) to look for a correlation between a data value and the ratio of a couple others and 2) to compare if jobs that take a long cycle time boards are easier or more difficult to predict than the shorter cycle time boards.
No comments:
Post a Comment