Case: enhancing statistical process control
Like most quality assurance systems, the customer's existing system was based on the prevailing thinking according to which variations in quality are to a major extent random and deviations must be accepted and that the trick is to afterwards find out the cause.
The customer wanted to enhance their process control. The customer had a rather old data collection system and a SPC (statistical process control) software installed in the quality organization. The system was developed over the years but the existing solution only considered some few variables simultaneously. It did not work satisfactorily either. For a process expert, this was revealed immediately by BayMiner's efficient visualization that initially did not show a natural clustering. The problem was aggravated by fact that the data collection focused on simple to collect but information poor data. Data about external factors was not collected, which made things worse. The existing solution provided sufficient functionality, as long as the process was simple and run by seasoned experts. However, with the increased use of new technology and inexperienced operators, the data was not enough to reveal the reasons to process disturbances.
The customer faced huge yield variations. The increased manufacturing and testing costs caused profitability problems and even worse, the customer had developed a classification system for symptoms that did not work. It led the process trouble-shooters in wrong directions, causing battle fatigue and poor discipline in the reporting.
Bayes Information Technology Ltd., a leading provider of predictive analytics using Bayesian modelling, was called in to solve the problem at low cost and on tight schedule. Bayes IT designed a combined knowledge model of the process parameters, external factors and available disturbance data. The rapidly created first version only verified the poor quality of the classification process. It gave however the customer's process experts indications, what more data should be collected. As a consequence, some data from the accounting dept. was included into the knowledge model to reveal correlations to logistics costs. In the second phase that took several months to realize, the data collection was further developed. Thanks to it, some additional process parameters could be included in the model. At the moment the variable definitions are clear and collected data of good quality.
The solution was designed so that it can be easily integrated to a SCM (Supply Chain Management) system. By using a de facto standard data format, BayMiner is in practice completely open (i.e. can be connected to any other system). There is no need for any special software, its updating, or expensive training.
The development continues; a risk model using BayMiner is under development with the objective to predict disturbances. Bayes Information Technology's genuine multi-variate analysis solution guarantees not only agility but also quality of the risk knowledge model. Cutting-edge technology ensures that prediction is accurate also in case data is missing.
With the process model designed by Bayes Information Technology, the customer was able to identify the reasons to their process disturbances. The new information enabled corrective actions to be made via a successful decision process. This would not have been possible using conventional statistical calculus of the data the company collected. Therefore, a knowledge model was created of the process.
The process downtime was reduced to less than 50 % of the level before the project. Several other yardsticks improved as well. The customer estimates that approximately 50 % of the overall positive impact can be attributed to the use of a knowledge model, the rest is due to many factors such as enhanced working environment.
The cost of the development project was approximately 100,000 euro, of which the major part related to data collection development. At the time of order, the payback period was estimated to be 1-2 years.
The objective of performing multivariate, probabilistic process control is to monitor the process over time in order to detect any unusual events enabling quality and process improvement. It is therefore essential to be able to track the cause of out of range observations since there may be several weak signals originating form the same disturbance source (co-occurrence). Opposed to uni-variate analysis that cannot reveal complex problem causes, the probabilistic Bayesian network models help to solve the complexity of multivariate problems and reveal not only the cross-correlation among even a great number of variables but also the frequently hidden co-occurences in these. When these relations are visualized properly, it is not difficult for the analyst to assign causes to the out-of-control signal.
The BayMiner method incorporates identifying of subgroups (clusters) and the analytical breakdown of the individual observations. Overlapping of the subgroups occur making the analysis work difficult. The BayMiner analysis technique enables the analyst to identify specific observations that occur in more than one subgroup. These must be identified and the reason either understood or a hypothesis formulated. With BayMiner this is easy.
There surely are various multi-variate methods (other than Bayminer) to choose from. Unfortunately, most of them use preset weights. Thus, a specific analysis does not reflect the reality in an environment where many essential factors change frequently. If the user, e.g., decides to give equal weight to every measured variable, the resulting visualization is directionally invariant, misleading the analyst.