Machine Learning and Artificial Intelligence for Process Optimization in Automated Fiber Placement of Composite Structures

Ply-drop design of non-conventional laminated composites using Bayesian optimization

The advent of Automated Fiber Placement (AFP) systems have aided the rapid manufacturing of composite aerospace structures. One of the challenges that AFP systems bring is the variability of the deposited prepreg tape layers, which are prone to gaps, overlaps and twists. The current detection method used in industry involves halting fabrication and performing a time consuming visual inspection of each tape layer. Typical AFP systems use a quartz lamp to heat the base layer to make the surface tacky as it deposits another tape layer. The innovation discussed in this paper is to use the preheated base layer as a through transmission heat source and inspect the newly added tape layer in situ using a thermographic camera mounted onto the AFP structure. Such a system would not only increase manufacturing throughput by reducing inspection times, but would also aid in process development for new structural designs or material systems. To this end, a small thermal camera was mounted onto an AFP robotic research platform at NASA, and thermal data was collected during typical and experimental layup operations. The data was post processed to reveal defects such as tow overlap/gap, wrinkling, and peel-up. Defects that would have been impossible to detect visually were also revealed in the data, such as poor/loss of adhesion between plies and the effects of vacuum debulking. This paper will cover the results of our experiments, and the recent progress on the data reduction algorithms in preparation for machine learning development.

Automated Fiber Placement (AFP) manufacturing methods, considered to be one of the most advanced methods of composite manufacturing, are used in fabricating complex and large scale composite structures. During AFP manufacturing, the gaps between each fiber-tow is closely controlled. However, due to machine/operational errors, gaps between these fiber-tows exceed set tolerances and are seen as a frequent feature/defect in AFP fabrication. By understanding the effect of these AFP induced gaps, we will be able to evaluate the integrity of the specimens/structures to better understand the effects due to these features/defects. Current research focuses on AFPgap configurations associated with missing tows and mainly looks into establishing methods for developing knockdowns/design allowable that can be incorporated into the design building block analysis and test validations.

Artificially intelligent process planner for automated fiber placement

Automated fibre placement (AFP) machines are able to place simultaneously several bundles of fibres, called tows, on a surface. Using AFP machines, it is also possible to manufacture composite laminates with fibres placed in curvilinear paths. The fibre orientations and stiffness properties of these laminates are spatially varied and hence they are called variable stiffness (VS) laminates in contrast to the traditional laminates with straight fibres which are called constant stiffness (CS) laminates. Past research has shown that the structural performance of laminated composite structures can be improved by spatial tailoring of laminate stiffness. One of the widely used structural components in aerospace industry are cylindrical shells. In this thesis, a computationally efficient framework was developed for circumferential laminate stiffness tailoring of unstiffened and longitudinally stiffened cylindrical shells with arbitrary cross sections to maximise the buckling capacity with consideration of strength constraints. In addition, the AFP manufacturing constraint on the maximum curvature of curvilinear fibre paths was considered. This constraint was applied to avoid wrinkling of the fibres placed inside the turn of a curved tow. The aforementioned framework utilised computationally efficient analysis and optimisation tools.

Residual deformation analysis in composite shell structures manufactured using automated fiber placement

Automated fibre placement (AFP) is a relatively new process for the manufacturing of composite structures. Among many attractive features, it provides high-speed of material deposition, more repeatability in terms of quality of the part, less labour intensive (as compared with traditional methods of manufacturing such as Hand Lay-Up), less waste and the ability to transition more seamlessly from design to manufacturing. AFP can be used to process both thermoset composites and thermoplastic composites. Thermoplastic composites processing holds many potential benefits. This is because if the process is done right in producing parts with good quality, it is fast since it does not require a second process such as curing in an autoclave or oven. For the purpose of comparison of performance and for design, it is necessary to determine the mechanical properties of laminates made using this process. However, there are challenges in making flat coupons for the purpose of testing for mechanical properties. This article presents these challenges and the procedure developed to make flat laminates using a simple AFP machine. Mechanical properties of these laminates are also determined and compared with those obtained from laminates made using conventional autoclave moulding.

Automated fibre placement (AFP) machine in-process layup inspection is essential for composite structure fabrication efficiency and quantity improvement. The laser profilometer scanning technique is one of the automated in-process inspection techniques. Laser profilometers can capture a large number of intensive high-resolution point clouds during the layup. To date, few studies have been published about processing point clouds at high speed for in-process layup defect detection and layup feature measurement. In this study, an algorithm called the cross-sectional line-processing algorithm was proposed, and a testbed was constructed to validate the algorithm. The algorithm processes each laser line captured by the laser profilometer and clusters the processed results together. Finally, the defect features and layup features can be segmented and recognized from the collected point cloud. Layup defect types can be recognized, and the dimensions of the layup features can be measured. The experimental results show that this approach can process at most 200 laser lines (about 160 000 points) per second, and the overall defect type recognition accuracy rate reaches about 78%, which meets basic AFP in-process inspection requirements.

Predictive computational modeling of the effect of automated fiber placement (AFP) defects on the structural performance of aerospace structural composites is of importance in preliminary design. A finite element method (FEM) based framework is presented for the analysis of the quasi-static tensile response of quasi-isotropic [45/0/ -?45/90]2s laminates containing controlled gaps and overlaps in the 90° plies with different defect sizes $\left (\frac {1}{2}", \frac {1}{4}", \frac {1}{8}", \frac {1}{16}"\right )$ . Consolidated ply geometries were obtained from micro-graphs of actual fabricated specimens in order to capture the effect of out-of-plane misalignments induced by these imperfections. The FEM models were generated in Abaqus CAE with a Python script, in a ply-by-ply manner. The material model for the virtual testing is a multiscale continuum progressive damage and failure model (MCDM), in which the prepeak nonlinear response is based on multiscale micromechanical formulations and the postpeak meso-scale response is modeled with the crack band method (CB). The FEM results were compared to experimental results and showed good agreement in failure modes and mechanical response. The ultimate tensile strength was predicted with reasonable accuracy for most of the cases. With both experimental and numerical analyses, this paper provides a better understanding of the progressive tensile failure of toughened composites (T800S/3900) with gap and overlap defects.

The functional advantages of tailored stiffness, often seen in nature, are also utilized in composite structures. Advancements in the multiaxis tow placement and automated fiber placement (AFP) machines led to the development of variable angle tow (VAT) composites, also referred further as variable stiffness composites (VSC). These composites are shown to effectively enhance the stress distribution and buckling load capacity of structures with greater flexibility on the design space. This review systematically presents the status of recent research on the topic of VSCs. Various manufacturing techniques of VSC are discussed; constraints and the defects associated with the manufacturing processes are enlisted. The review highlights the optimization studies based on the fiber profile and macro-scale stiffness invariants. Several studies existing in the domain of buckling, vibration, and aeroelastic tailoring of angle tow composites are summarized to connect the important aspects of analysis and present a holistic approach for future studies in this area.

A ply surface modeling and path planning approach is presented to reduce placement defects in complex composite structures with large and variable thickness. The mechanisms of thickness‐induced defects are analyzed through mathematical modeling. An algorithm based on the fast marching method (FMM) is developed to generate ply surfaces for complex‐geometry components with large and variable thickness. This ply surface modeling is integrated with a mesh‐based path planning algorithm designed to minimize defects on thick and variable‐thickness structures. Simulation and experimental results confirm significant improvements: automated fiber placement (AFP) trials on a spar show that 45° ply gaps decrease from 3.21 to 1.45 mm, satisfying tolerance requirements, and all wrinkle defects are eliminated. This work provides an effective solution for low‐defect AFP manufacturing of complex, thick composite components.

Manufacturable insight into modelling and design considerations in fibre-steered composite laminates: State of the art and perspective

Topology and fiber path optimizations result in the ideal design of the shape and fiber placement, respectively, in composite structures subjected to certain loads and constraints. Advancements in additive manufacturing (AM) can enable such designs to be manufactured. Specifically, the use of continuous carbon fiber-reinforced plastics (CFRPs) in AM or automated fiber placement (AFP) allows for 3D printing of such topology and fiber path optimized designs. However, fiber path optimization methods generally result in a collection of fibers in loops (closed loops) that are impossible to print, and lines (open loops), which may be too short to print depending on the system. Additionally, placement of discontinuities, i.e., where a fiber path ends inside a part, can result in stress concentrations that degrade the part performance. This paper introduces an approach to automatically convert the collection of closed and open loops into printable set(s) of fibers. An additional optional consideration of stress concenterations can be provided to the algorithm, which is demonstrated using finite element analysis (FEA) for enhanced structural performance.

We are witnessing a transformation in manufacturing industries driven by integrating Industry 4.0 infrastructures and smart manufacturing systems. This is being made possible by leveraging advanced analytical technologies like AI, Machine Learning (ML), and Digital Twins leading to lower costs, higher efficiency, and quality improvements. Time-series analytics plays a pivotal role in the manufacturing domain due to the abundance of time-series, and it has the potential to be used in different applications such as fault detection, predictive maintenance, energy and demand forecasting, and remaining useful life prediction. Consequently, Time-Series Classification (TSC) and Time-Series Forecasting (TSF) are two main tasks in this domain to set the foundation for predictive applications in smart manufacturing systems. Over the past decade, researchers have introduced many algorithms for TSC and TSF, necessitating validation, and empirical comparison of them alongside development. This dual approach holds substantial value for researchers and practitioners by streamlining choices and revealing insights into models' strengths and weaknesses. There is a gap in manufacturing where theoretical advancements have not been fully translated into practical applications. The main objective of this Ph.D. dissertation is to develop a comprehensive framework that will help to bridge the gap, provide guidance for practitioners, and advance data analytics in manufacturing systems. This was done by performing a rigorous experimental evaluation of SoTA ML and DL algorithms on a variety of datasets for the mentioned tasks. Additionally, the scarcity of manufacturing datasets was addressed by providing sets of manufacturing-related datasets, preprocessed to facilitate algorithm implementation. For the TSC framework, 92 algorithms were identified, reviewed, and categorized. These included ML, ANN, and DL algorithms from different feature engineering and classification techniques. Thirty-six most representative algorithms were methodologically selected for evaluation on a set of 22 manufacturing datasets with different characteristics, covering diverse manufacturing problems. For TSF, more than 100 algorithms from Statistical, ML, and DL approaches and 29 datasets were studied. Subsequently, the experimental evaluation was conducted on a subset of 18 representative algorithms and 13 datasets over four different experimental scenarios. Subtopics and use cases for TSC and TSF frameworks were also explored, including high-resolution time-series classification and real-time defect detection in a robotic assembly line using original research data. By conducting extensive evaluations, creating curated datasets, and demonstrating real-world applications, this dissertation aims to contribute to the body of knowledge for more efficient data-driven decision-making in manufacturing industries. The developed frameworks are an effort toward this goal

In the dynamic landscape of advanced manufacturing, the confluence of laser powder bed fusion (LPBF) and machine learning (ML) has recently garnered significant attention in many applications. This review investigates the confluence of LPBF and ML, specifically within the specific domain of stainless steel. Firstly, it delves into LPBF principles, including an overview of critical process parameters and associated defects. Secondly, the paper meticulously addresses the distinct challenges posed by steel in additive manufacturing (AM), highlighting factors such as chemical composition, anisotropic microstructure, and oxide film formation, all of which require specialized considerations. Thirdly, the spotlight shifts to the pivotal role of ML, covering predictive modeling for process parameters, real-time defect detection, and quality control. This paper highlights recent advances, revealing how data-driven approaches can accelerate process understanding and part qualification. Eventually, this review offers insights into the future integration of ML in LPBF for steel, providing valuable perspectives on potential advancements in the field of AM.