Metamodeling has been widely used in engineering for simplifying predictions of behavior in complex systems. The kriging method (Gaussian Process Regression) could be considered as a metamodeling technique that uses spatial correlations of sampling points to predict outcomes in complex and random processes. However, for large and nonideal data sets typical to those found in complex manufacturing scenarios, the kriging method is susceptible to losing its predictability and efficiency. To address these potential vulnerabilities, this article introduces a novel, dynamic metamodeling method that adapts kriging covariance matrices to improve predictability in contextualized, nonideal data sets. A key highlight of this approach is the optimal linking process, based on the location of prospective points, to alter the conventional stationary covariance matrices. This process reduces the size of resulting dynamic covariance matrices by retaining only the most critical elements necessary to maintain accuracy and reliability of new-point predictability. To further improve model fidelity, both the Gaussian parameters and design space attributes are optimized holistically within a problem space. Case studies with a representative test function show that the resulting Dynamic Variance-Covariance Matrix (DVCM) method is highly efficient without compromising accuracy. A second case study representative of an advanced manufacturing setting demonstrates the applicability and advantages of the DVCM method, including significantly increased model robustness.

Additive manufacturing (AM) is a new and disruptive technology that comes with a set of unique challenges. One of them is the lack of understanding of the complex relationships between the numerous physical phenomena occurring in these processes. Metamodels can be used to provide a simplified mathematical framework for capturing the behavior of such complex systems. At the same time, they offer a reusable and composable paradigm to study, analyze, diagnose, forecast, and design AM parts and process plans. Training a metamodel requires a large number of experiments and even more so in AM due to the various process parameters involved. To address this challenge, this work analyzes and prescribes metamodeling techniques to select optimal sample points, construct and update metamodels, and test them for specific and isolated physical phenomena. A simplified case study of two different laser welding process experiments is presented to illustrate the potential use of these concepts. We conclude with a discussion on potential future directions, such as data and model integration while also accounting for sources of uncertainty.

Many complex electromechanical assemblies that are essential to the vital function of certain products can be time-consuming to inspect to a sufficient level of certainty. Examples include subsystems of machine tools, robots, aircraft, and automobiles. Out-of-tolerance conditions can occur due to either random common-cause variability or undetected nonstandard deviations, such as those posed by debris from foreign objects. New methods need to be implemented to enable the utilization of detection technologies in ways that can significantly reduce inspection efforts. Some of the most informative three-dimensional image recognition methods may not be sufficiently reliable or versatile enough for a wide diversity of assemblies. It can also be an extensive process to train the recognition on all possible anomalies comprehensively enough for inspection certainty. This paper introduces a methodical technique to implement a semiautonomous inspection system and its algorithm, introduced in a prior publication, that can learn manufacturing inspection inference from image recognition capabilities. This fundamental capability accepts data inputs that can be obtained during the image recognition training process followed by machine learning of the likely results. The resulting intelligent insights can inform an inspector of the likelihood that an assembly scanned by image recognition technology will meet the manufacturing specifications. An experimental design is introduced to generate data that can train and test models with a realistic representation of manufacturing cases. A benchmark case study example is presented to enable comparison to models from manufacturing cases. The fundamental method is demonstrated using a realistic assembly manufacturing example. Recommendations are given to guide efforts to deploy this entire methodical technique comprehensively.

As the life cycles of many products advance, obsolescence of original components can pose challenges for a product’s maintainability and serviceability. Information about the original design intent of the product system may not be available. For example, it may not be obvious how to identify old components by reverse engineering or to select and configure the optimal components for best system performance. Thus, time spent to redesign an older legacy product with high-fidelity simulations may be difficult to justify. Increasingly more performance measurements are available in digital twins of product systems generated during manufacturing and testing. This paper introduces a redesign from manufacturing data mapping (RfMDM) method that can infer optimal new components to modernize legacy product systems. This RfMDM method does not rely upon time-consuming redesign efforts. The inputs come from a basic understanding of the subsystem of concern and qualitative scoring estimates of the influence of design rules on design criteria and simulation fidelity. RfMDM transforms manufacturing data to determine the significance and required quality of each design variable of each obsolete component. This enables estimation of optimal new component specifications for selection and a first article redesign and test to enable product line improvements over time more efficiently. Recognizing that radiofrequency interference effects in printed circuit boards present a complex and quintessential example of legacy subsystems with potentially obsolete components, this paper demonstrates the RfMDM method with a theoretical example subsystem of an active low-pass filter on a printed circuit board.

Final test metrics that evaluate product system performance usually depend upon numerous variables, such as dimensions or other characteristics of parts and assemblies. Many product systems are expensively comprised of numerous parts. Therefore, during new product system development, the challenge becomes how to rapidly learn estimated system results from combinations of many variables at the smallest possible sample size to minimize cost and improve product quality. In this work, we introduce a fundamental Vector-Based Clustering technique to predict a cluster range of system test results for comparison to other machine learning techniques in a commercial software tool. This work expands to include two additional techniques that account for significance among many variables. All three of these techniques were tested and compared to the machine learning algorithm from a commercial tool best suited for each training set from a high dimensional open-source data set representative of manufacturing system data. These case study results show improvement in predictive accuracy over many prevalent machine learning techniques at small sample sizes. Furthermore, since a best-suited machine learning technique is selected by trial and error for each training set, the computational time is significantly improved as well.