Uncertainty Quantification Based on Bayesian Neural Networks for Predictive Quality

In the context of production metrology, the field Predictive Quality develops methods based on statistics and machine learning to predict quality characteristics from process data. In prior work, conventional machine learning methods such as feed-forward neural networks have been successfully applied. Yet, an uncertainty quantification for the prediction is not provided. Therefore, it is not possible to prove the suitability of the applied predictive quality methods for quality inspections. However, we can estimate the uncertainty by taking a Bayesian perspective and utilizing suitable algorithms. Here we define Prediction of Quality Characteristics (PQC), which is the foundation for every Predictive Quality application. We extend our definition of PQC into a general Bayesian framework to interpret predicted quality characteristics. As an example, we show how Bayesian neural networks are applied to PQC to estimate the uncertainty of every prediction. We interpret the results in the industrial context and determine the suitability of the PQC method. Our results demonstrate that the application of Bayesian methods is highly promising to get Predictive Quality recognized in industry as an accredited method for quality inspections.