To recognize in the production line defected or troubling production units in time reduces equipment down times, scrap and expensive customer returns, which are likely to happen e.g. for semiconductors in the automotive industry. PMC is leading new ways to use machine learning methods as predictors. These methods are adaptive to new fields of applications (e.g. new equipments or technologies) by training them with respective production data.
PMC is a pioneer in the use of advanced machine learning for Predictive Maintenance and Virtual Metrology. Therefore first the knowledge of production engineers is evaluated and formalized. On the basis of this preparatory step, a statistical model can then be set up. (e.g. testing, information theory, unsupervised learning methods).
Especially for the production disruption significant variables can be identified (feature selection) which can help to find the reason for the down times. After the data has been subject to preliminary filtering, we use adaptive statistical predictors, which are robust against fault measuring, limited data and differing production context (support vector machine, ridge regression, time series analysis).
We are evaluating the success using industrial standards (e.g. Gauge Repeatability) or we use expertise knowledge to create cost functions that allows minimizing the cost of production in the specific context.