Data Intensive Method For Processing Defect Detection and Mitigation For Composites

Composite manufacturing without processing defects is a crucial step in satisfying the production, performance, and quality requirements of composite materials in various industries. Autoclave processing enables manufacturing of high-quality composite parts with excellent mechanical properties, whereas additive manufacturing (AM) offers adaptability to complex designs and streamlined processes. However, each method presents unique challenges; despite the benefits, autoclave processes can still result in processing defects, and AM is particularly susceptible to processing anomalies owing to the novelty of the process. Ensuring the quality and reliability of composite manufacturing is essential to fully capitalize on the advantages offered by both techniques. Artificial intelligence (AI) applications in composite manufacturing offer promising solutions for automated quality inspection; however, practical implementation is limited by the lack of in situ imaging data and effective anomaly detection models trained specifically for defect detection during composite curing. In this study, different methods were proposed to detect and reduce processing defects in conventional and additive manufacturing of composite materials. For a conventional autoclave, a method was proposed to reduce the residual stress by modifying the thermomechanical properties at the fiber-matrix interface. A novel digital image correlation (DIC)-based in situ approach was utilized to evaluate the strains and residual stresses during autoclave curing of the composites. Interface modification results in a reduced residual stress and increased laminate strength and stiffness. For AI application in composite processing, a novel dataset of the composite curing process was developed and utilized to develop an explainable AI framework for real-time defect detection using a zero-bias deep neural network (ZBDNN). In this method, the last dense layer of the deep neural network (DNN) is replaced by two consecutive parts: a regular dense layer denoted (L1) for dimensional reduction and a similarity matching layer (L2) for equal weight and non-biased cosine similarity matching. Subsequently, a new three-step anomaly detection algorithm was developed using an autoencoder, unsupervised machine learning, and an elliptical envelope mathematical model to reduce the retraining and improve the detection timing performance. The anomaly detection performance of both the models was evaluated using an autoclaved dataset. Subsequently, the ZBDNN was applied to detect defects during fused deposition modeling (FDM) of thermoplastic polymers, and multi source data were generated using a multi-camera system, including thermal and charge-coupled device (CCD) cameras for composite 3D printing. It was demonstrated that the approach is capable of successfully detecting multiple types of defects, such as cracks, stringing, and warping, with high accuracy. It was also concluded that for anomaly detection, the ZBDNN outperformed the one-class support vector machine and autoencoder, and the use of multi source data improved the anomaly detection accuracy.