A novel approach for quality control of automated production lines working under highly inconsistent conditions.

When addressing product quality standards in manufacturing lines, a critical issue is the identification of the parameters that define the quality of the final product and their tracking. The problem of process control under inconsistent working condition of an automatic machinery, i.e. when some parameters are highly variable, is still quite unexplored in literature. This objective becomes even more challenging when the most important process variables are not directly measurable. This paper demonstrates that it is possible to achieve quality control by coupling a soft sensor, whose predictive model is a neural network, with an anomaly detector. The methodology has been applied to automatic machinery placed in a manufacturing line, where high variability in production rate has an important effect on the measured physical variables. This makes automated and accurate quality control difficult, due to the fact that in this test case the data collected are accelerometers signals, extremely sensible to variation in machine productivity by definition. It is shown that this approach outperforms many other classification methods (Support Vector Machines, Ensemble Bagged Tree, Discriminant Analysis, K-nearest neighbours and the direct application of a Neural Network) proposed in the past, achieving satisfactory results evaluated on the basis of four metrics (Accuracy, precision, recall and F 1 -score), even if anomalous data have been collected in a limited number of machine's working points. In particular, an accuracy over 92% has been reached also for production rates where only nominal conditions are collected. This procedure exceeds the direct training of a neural network (accuracy of 57.6% at new production rates), as well as the application of shallow methods based on the extraction of dimensionless features (around 35% in accuracy at new production rates). • Quality control is arduous when it is not possible to measure process variables. • Hard to build robust soft sensor with system working under inconsistent conditions. • A neural network can be trained at specific machine working points. • Anomaly detection is achieved by deep learning and a probabilistic model. • Accuracy higher than 92% for each production rate tested. [ABSTRACT FROM AUTHOR]