Methods for Quality Monitoring in Ultrasonic Welding of Carbon Fiber Reinforced Polymer Composites

Abstract

Carbon fiber reinforced composites have been increasingly used in various industrial sectors, especially in the automotive industry. Ultrasonic welding is considered as an effective approach to joining such composites. Reliable weld quality classification and prediction methods are needed to ensure quality and reduce manufacturing costs. However, existing methods have two weaknesses. The first one is that the majority of the existing methods are based on signal feature data extracted from the original experimental time-series data. Feature-based models may not take full advantage of the information contained in the large amounts of time-series data available, even though the models are simple and easy to program. On the other hand, when using experimental time-series data to conduct weld quality monitoring, the data size may be insufficient for training neural network-based methods for quality monitoring or classification. Therefore, a method is needed to augment experimental data while preserving the statistical characteristics of the experimental data. To find reliable quality monitoring models in various situations, this dissertation proposes two neural network models that are respectively applied to feature-based data and full time-series-based data and compares their performances. The dissertation first investigates the relationship between weld energy and joint performance in ultrasonic welding of carbon fiber reinforced polymer (CFRP) sheets through weld experiments. The weld quality classes for training quality monitoring algorithms are determined from welded joint lap-shear strength and the microstructure of the weld zone. These pre-defined weld quality classes are the output criteria for weld quality monitoring on feature-based models and time-series-based models. For feature- based weld quality monitoring, a simple and efficient feature selection method is first developed to screen the most significant features for classification from multiple weld quality classes. A Bayesian regularized neural network (BRNN) is then demonstrated to be more accurate and robust when classifying weld quality classes in ultrasonic composite welding when using feature-based data as the input than the previously proposed methods of support vector machine (SVM), k-nearest neighbors (kNN), and linear discriminant analysis (LDA). To address the limited size of experimental data, a Multivariate Monte Carlo (MMC) simulation with copulas approach is proposed to reasonably generate large amounts of time-series process signals for ultrasonic composite welding. With both experimental data and a large quantity of simulated data, a deep convolutional neural network (CNN) is applied to weld quality classification. The CNN model is found to be more accurate and robust, not only under small training data set sizes, but also under large training data set sizes when compared with previously researched classification methods applied in ultrasonic welding. In conclusion, neural network-based models could achieve high accuracy using feature signals and the full time-series process signals.