Sensing Uncertainty Evaluation for Product Quality

Abstract

As sensing techniques continue to enhance the manufacturing industry in terms of condition monitoring, process control and decision making, understanding uncertainty involved in sensing has remained a challenging problem. Sensor readings may be affected by deterioration of the hardware and environmental perturbations. Prior studies reported in the literature assumed multiple co-located homogeneous sensors to provide redundant information for uncertainty evaluation. Such an approach is often times not applicable given space restraint and cost-effectiveness concerns. Furthermore, it is difficult to distinguish the sources that have caused the variation of sensor readings, e.g., limitation in the sensor precision or variation in the measured quantity. To address these challenges, a method for computing uncertainty of non-homogeneous sensors is developed, using on-line injection molding quality monitoring as experimental verification. The method includes two computational steps. First, measurements provided by four spatially distributed sensors (two temperature and two pressure sensors) in an injection mould cavity are fused to estimate the part quality (thickness). Second, the ground truth is approximated from the estimated part quality, through inverse process. Subsequently, the approximated ground truth is utilized to calculate the uncertainty of each sensor used in the measurement process. To quantify sensor uncertainty, a metric including accuracy, precision and trust is defined. The detection of abnormal sensors in turn provides input to improved part quality estimation. The developed technique is evaluated experimentally by measurements on a production-grade injection molding machine working under a variety of machining settings. The results demonstrate the approach's effectiveness for evaluating sensing uncertainty and improving monitoring performance.