CORRELATIONS IN MANUFACTURING PROCESS DATA ANALYSIS – Part 2
Esa Vilkama President and Founder at Process Data Insights, LLC
Design of Experiments
In industrial experimentation, the Design of Experiments (DOE) is usually the best and fastest way to learn cause-effect correlations between independent (factors, inputs) and dependent (responses, outputs) process variables. It allows for multiple factors to be systematically and simultaneously changed, determining their effect on a desired response. DOE is faster and better than the one-factor-at-a-time method (OFAT) of changing values, and also finds interactions that OFAT misses. DOE can be used to reduce product design and manufacturing costs by finding the optimal combination of values for the input factors.
There are many types of DOE methods. A Screening Design is to identify the most important factors that affect process quality. A Full Factorial is an experiment whose design consists of two or more factors, each with discrete possible values or “levels”, and using all possible combinations of these levels across all factors. Such an experiment allows the experimenter to quickly study the effect of each factor on the response variable, as well as the effects of interactions between factors on the response variable.
If it is not possible use all the combinations of the factor levels, Fractional Factorial designs can be used to lower the cost and speed up experimentation. However, some effects are then confounded (not distinguishable) with each other. Response Surface Methodology (RSM) is an effective method to optimize process conditions, and it can be used to fit a complete quadratic polynomial model. Mixture experiments are a special class of response surface experiments in which the product under investigation is made up of several components or ingredients.
As a drawback, DOEs are often impossible or impractical to do because they interfere with or interrupt normal production, or are otherwise difficult to perform. For example, for a factorial design, some combinations of factor values may be unsafe or impossible to use. Also, running an experiment can be expensive on pilot or production lines because of raw materials, energy and labor needed.
Commercial: Minitab® (https://www.minitab.com), Design-Expert® (https://www.statease.com/software/design-expert/), JMP® software from SAS (https://www.jmp.com), MATLAB® Statistics and Machine Learning Toolbox™ (https://www.mathworks.com/products/statistics.html), Isight (https://www.3ds.com/products-services/simulia/products/isight-simulia-execution-engine/), QI Macros (https://www.qimacros.com/), Aexd.net (https://aexd.net/), ECHIP (http://www.echip.com/), Quantum XL (http://sigmazone.com/quantumxl_features/), Genstats (https://www.vsni.co.uk/software/genstat)
Open-source: Design of Experiments in R (https://cran.r-project.org/web/views/ExperimentalDesign.html), Develve (http://develve.net/), pyDOE: The experimental design package for Python (https://pythonhosted.org/pyDOE/), Expyriment (https://www.expyriment.org/)
Evolutionary Operation (EVOP) is a statistical experimentation technique to introduce a carefully planned repeated pattern of small changes in the manufacturing process variables, within specifications, during normal production without interfering with or interrupting it. These changes are not large enough to result in non-conforming product, but are significant enough to detect the effects of the adjustments which are then evaluated just as with DOE. The process is then shifted in the desired direction of improvement. EVOP is based on the condition that normal production has the ability to contribute valuable information on the effect (correlation) of process variables on a particular product characteristic or feature. The resulting data contains information for continuous process improvement that can also be used for machine learning.
Since small incremental changes in process parameters, Evolutionary Operation does not require any special resources. EVOP is not a substitute for the fundamental investigation and experimentation. EVOP can complement DOE, doing “fine-tuning” of the process after startup, e.g. when going from the pilot plant to the full-scale production mode. EVOP is an iterative type of DOE that will help you find the optimized solution. EVOP has been successfully used in many manufacturing applications to improve the performance of industrial processes.
In the next blog I will cover how to use Data Visualization in finding correlations in manufacturing data.
Please contact us at Process Data Insights, LLC, if you are interested in a comprehensive correlation analysis of your manufacturing process.