Coaxial cable manufacturing process and forecasting model
Process control is one of the most important parts in high quality coaxial cable manufacturing. Well known statistical process control techniques are very inflexible, the models must be updated manually at any process change. New techniques with self learning algorithms are need to control such a dynamic and high dimensional system like a coaxial cable production. There's also a request for forecasting product values during manufacturing and prediction of untested product parts.
To build such a forecasting system, there must be a system to acquire all process relevant values, like pressures, diameters, electrical values (insertion/return loss, phase stability, etc.), rotation speed, temperatures, material characteristics, etc. All this data must be stored and purified for next analytical and forecasting steps.
Forecasting quality increases all the more results of next process steps are available. So at the end, forecasting over each untested part of production can be done.
In this thesis a generic DAQ infrastructure was implemented, with capabilities for future extensions and a generic interface for different forecasting models implemented in .Net and Matlab. As one production takes over 6 weeks from the first step to the end, it was not possible to forecast the cable end characteristics. So this thesis is limited to inline production and algorithm implementation.
The task of data acquiring was a simple software engineering task with focus on performance and extensibility. The bigger part of this thesis were algorithm studying and implementation. It's been shown, that machine learning algorithm are very simple, if they are understood.
So inline forecasting was implemented and refined in the PTFE dielectric production step. To forecast the center conductor eccentricity. This process step is not as complex, as there are only about seven parameters that have an impact on eccentricity. One of the biggest knowledge was, that synthetic inputs and preprocessing are very important for good forecasting results. In this section there is much work to do, to refine forecasting results. On of the biggest awareness was, that algorithm are just so good, as there training data.
For future improvement of the system to a productive status, there are more synthetic inputs, new synthetic inputs and more data needed. There is also a test environment needed, to test new implementations of algorithms based on real data in short time. This will help improve failure detection in algorithm and filter, which results in a much better forecasting result.