Better quality management in manufacturing systems usually means preventing defects, reducing carbon emissions, and making systems greener. Modeling stress-induced processes is challenging and extremely critical in the quality management of advanced manufacturing systems. While residual stresses may be beneficial in some situations, in composite aircraft assembly, high residual stresses and extreme deformations are crucial and must be accounted for to prevent future catastrophic failures. Currently, conventional approaches to the optimal placement of actuators on composite structures are non-optimal, require ten actuators heuristically, and do not consider residual stresses. To overcome these limitations, we propose a Stress-Aware Optimal Actuator Placement framework. We provide theoretical investigations that demonstrate the convergence to global optimum, computational complexity, and mean prediction error of the proposed optimization algorithm. The stress-aware optimal actuator placement framework is able to achieve significant reductions of at least 39.3% in mean root mean squared deviations (RMSD) and 52% in maximum forces (MF), and only requires eight actuators on average while satisfying the safety threshold of residual stresses. Note to Practitioners—In aerospace manufacturing, about 80% of defects are associated with the assembly process. Defects may result in large errors and waste, high energy costs, low product quality, or even endanger human lives. The actuator placement usually significantly impacts the final quality of the assembled airplanes. Existing actuator placement strategies are not sufficient for composite aircraft assembly due to the complex nonlinear properties, ultra-high precision requirement, and residual stress requirement. The proposed method can improve the dimensional quality by optimizing actuator placement, as well as lower the residual stress and ensure the safety of products. Although our optimization framework was applied to the placement of actuators on composite fuselages, it could also be applied to the engineering design of other actuating systems in which both dimensional quality and stresses are sensitive. The proposed approach can reduce carbon emissions by preventing defects and improving quality management, ultimately aiming at zero-defect green manufacturing.

Real-time production monitoring incorporating progress prediction and anomaly detection is essential for quality and efficiency. Traditional vision-based anomaly detection methods struggle to differentiate between production-related features and background noise, and fail to consider the heterogeneity of production stages. This paper introduces an integrated approach that merges progress prediction and anomaly detection, employing the Autoencoder Process Probability Embedding (APPE) method. APPE maps the distribution of images from normal production to a progress-related Gaussian Mixture Model (GMM), focusing on identifying production-relevant features while minimizing background interference through the proposed Spatial Activation Map (SAM). The proposed SAM improves the interpretability of the neural network by highlighting the specific features that influence the model’s decisions. The method is assessed through real-world datasets in the assembly of water valves and the production of commercial aircraft spoilers. The case study shows that our approach can achieve superior effectiveness compared to the benchmark, notably improving both task performances by integrating progress prediction with anomaly detection. Note to Practitioners—In many manufacturing settings, such as aircraft production, tasks that involve human-robot collaboration or high-precision manual assembly play a significant role. The ability to detect anomalies and monitor progress in real-time is critical for ensuring the quality and efficiency of production. The manual nature of these operations makes them challenging to monitor through in-situ embedded digital sensors, yet real-time operation videos are readily available. Vision-based production monitoring has been widely used in applications such as product surface inspection, but existing algorithms often face difficulties distinguishing between normal background variations and anomalies related to production. This paper introduces a new approach, called Autoencoder Process Probability Embedding (APPE), which integrates progress recognition and anomaly detection into a cohesive monitoring task, allowing the model to differentiate between background elements and features related to production. Although our method is demonstrated in production scenarios as case studies, the proposed SAM mechanism is versatile to be applied in other contexts with similar types of categorical labels.

The training and testing data for deep-neural-network-based classifiers are usually assumed to be sampled from the same distribution. When part of the testing samples are drawn from a distribution that is sufficiently far away from that of the training samples (a.k.a. out-of-distribution (OOD) samples), the trained neural network has a tendency to make high-confidence predictions for these OOD samples. Detection of the OOD samples is critical when training a neural network used for image classification, object detection, etc. It can enhance the classifier's robustness to irrelevant inputs, and improve the system's resilience and security under different forms of attacks. Detection of OOD samples has three main challenges: (i) the proposed OOD detection method should be compatible with various architectures of classifiers (e.g., DenseNet, ResNet) without significantly increasing the model complexity and requirements on computational resources; (ii) the OOD samples may come from multiple distributions, whose class labels are commonly unavailable; (iii) a score function needs to be defined to effectively separate OOD samples from in-distribution (InD) samples. To overcome these challenges, we propose a Wasserstein-based out-of-distribution detection (WOOD) method. The basic idea is to define a Wasserstein-based score that evaluates the dissimilarity between a test sample and the distribution of InD samples. An optimization problem is then formulated and solved based on the proposed score function. The statistical learning bound of the proposed method is investigated to guarantee that the loss value achieved by the empirical optimizer approximates the global optimum. The comparison study results demonstrate that the proposed WOOD consistently outperforms other existing OOD detection methods.

Complex constrained global optimization problems such as optimal actuators placement are extremely challenging. Such challenges, including nonlinearity and nonstationarity of engineering response surfaces, hinder the use of ordinary constrained Bayesian optimization (CBO) techniques with standard Gaussian processes as surrogate models. To overcome those challenges, we propose a physics-constrained Bayesian optimization with multi-layer deep structured Gaussian processes, MGP-CBO. Specifically, we introduce a surrogate model with a multi-layer deep Gaussian process (MGP) mean function. The hierarchical structure of our model enables the applicability of constrained Bayesian optimization to complex nonlinear and nonstationary processes. The deep Gaussian process regression model, MGP, can efficiently and effectively represent the response surface function between actuators and dimensional deformations, thus yielding a better estimated global optimum in a shorter computational time. The proposed MGP-CBO model can realize faster convergence to the global optimum with lower constraint violations. Through extensive evaluations carried out on synthetic problems and a real-world engineering design problem, we show that MGP-CBO outperforms existing benchmarks. Although we use the optimal actuators placement as a demonstration example, the proposed MGP-CBO model can be applied to other complex nonstationary engineering optimization problems. Note to Practitioners—Bayesian optimization is a widely used sequential design strategy for engineering optimization because it does not rely on functional forms of response surfaces. This paper helps address two questions in practice: (i) how to incorporate physics constraints into Bayesian optimization. (ii) How to do Bayesian optimization when the systems have hierarchical structures. In practice, the hierarchical system structure is ubiquitous, and the engineering optimization is constrained by physical laws or special requirements. Therefore, the proposed physics-constrained Bayesian optimization with a multi-layer Gaussian process could provide a new tool for engineering design optimization problems. The computational convergence and complexity have been investigated. The proposed method is applicable to broad complex and nonstationary engineering optimization problems.

Active learning is a subfield of machine learning that is devised for the design and modeling of systems with highly expensive sampling costs. Industrial and engineering systems are generally subject to physics constraints that may induce fatal failures when they are violated, while such constraints are frequently underestimated in active learning. In this paper, we develop a novel active learning method that avoids failures considering implicit physics constraints that govern the system. The proposed approach is driven by two tasks: safe variance reduction explores the safe region to reduce the variance of the target model, and safe region expansion aims to extend the explorable region. The integrated acquisition function is devised to conflate two tasks and judiciously optimize them. The proposed method is applied to the composite fuselage assembly process with consideration of material failure using the Tsai-Wu criterion, and it is able to achieve zero failure without the knowledge of explicit failure regions. Note to Practitioners—This paper is motivated by engineering systems with implicit physics constraints related to system failures. Implicit physics constraints refer to failure processes in which explicit analytic forms do not exist, so demanding numerical simulations or real experiments are required to check one’s safety. The main objective of this paper is to develop an active learning strategy that safely learns the target process in the system by minimizing failures without preliminary reliability analysis. The proposed method mainly targets real systems whose failure conditions are not thoroughly investigated or uncertain. We applied the proposed method to the predictive modeling of composite fuselage deformation in the aircraft manufacturing process, and it achieved zero failure in sampling by considering the composite failure criterion.

Some response surface functions in complex engineering systems are usually highly nonlinear, unformed, and expensive to evaluate. To tackle this challenge, Bayesian optimization (BO), which conducts sequential design via a posterior distribution over the objective function, is a critical method used to find the global optimum of black-box functions. Kernel functions play an important role in shaping the posterior distribution of the estimated function. The widely used kernel function, e.g., radial basis function (RBF), is very vulnerable and susceptible to outliers; the existence of outliers is causing its Gaussian process (GP) surrogate model to be sporadic. In this article, we propose a robust kernel function, asymmetric elastic net radial basis function (AEN-RBF). Its validity as a kernel function and computational complexity are evaluated. When compared with the baseline RBF kernel, we prove theoretically that AEN-RBF can realize smaller mean squared prediction error under mild conditions. The proposed AEN-RBF kernel function can also realize faster convergence to the global optimum. We also show that the AEN-RBF kernel function is less sensitive to outliers, and hence improves the robustness of the corresponding BO with GPs. Through extensive evaluations carried out on synthetic and real-world optimization problems, we show that AEN-RBF outperforms the existing benchmark kernel functions. Note to Practitioners—Some industrial systems cannot be accurately represented by physical models. In this situation, data-driven black-box optimization is necessary for advancing the system automation and intelligence. BO is one of the widely used strategies for learning the global optimum of black-box functions. BO has been applied to robotics, anomaly detection, automatic learning algorithm configuration, reinforcement learning, and deep learning. This article proposes one new kernel function, named after AEN-RBF. The new kernel function will make BO with GPs more robust to outliers and lower the data quality barrier of model training. This article was motivated by the hyperparameter tuning problem of deep learning models for image defect detection in advanced manufacturing, but the method can be easily extended to other applications where kernel functions are needed. Our proposed method is verified by synthetic and real-world optimization problems.

This article develops a novel in situ nondestructive quality evaluation for friction stir blind riveting in joining lightweight materials. This method is able to solve the small sample size problem that is commonly occurred in manufacturing experiments. The proposed method achieves an optimal integration of the tensor decomposition and ensemble learning by utilizing the mutual benefits. On the one hand, diversified feature matrices are extracted via tensor decomposition to maximize the ensemble learning performance. On the other hand, regularized tensor decomposition results deviate with different regularization parameter values and ensemble learning is able to determine the optimal parameter value via a heuristic algorithm, which stabilizes the tensor decomposition results. This optimal integration is built by developing a novel diversity-based feature generation and selection approach: 1) a diversity measure is defined to evaluate the extracted features; 2) a heuristic adaptive algorithm is developed with ensemble learning to determine the optimal regularization parameter for integration; and 3) the optimal features are selected via clustering to maximize the diversity measure, which is expected to strengthen ensemble learning performance for better evaluation results. Numerical studies and case studies are performed to demonstrate the effectiveness of the proposed method as well as its superiority over the existing methods.

In the machine learning domain, active learning is an iterative data selection algorithm for maximizing information acquisition and improving model performance with limited training samples. It is very useful, especially for industrial applications where training samples are expensive, time-consuming, or difficult to obtain. Existing methods mainly focus on active learning for classification, and a few methods are designed for regression, such as linear regression or Gaussian process (GP). Uncertainties from measurement errors and intrinsic input noise inevitably exist in the experimental data, which further affects the modeling performance. The existing active learning methods do not incorporate these uncertainties for GP. In this article, we propose two new active learning algorithms for the GP with uncertainties, which are variance-based weighted active learning algorithm and D-optimal weighted active learning algorithm. Through numerical study, we show that the proposed approach can incorporate the impact of uncertainties and realize better prediction performance. This approach has been applied to improving the predictive modeling for automatic shape control of composite fuselage.

With the rapid development of sensor technologies, multisensor signals are now readily available for health condition monitoring and remaining useful life (RUL) prediction. To fully utilize these signals for a better health condition assessment and RUL prediction, health indices are often constructed through various data fusion techniques. Nevertheless, most of the existing methods fuse signals linearly, which may not be sufficient to characterize the health status for RUL prediction. To address this issue and improve the predictability, this article proposes a novel nonlinear data fusion approach, namely, a shape-constrained neural data fusion network for health index construction. Especially, a neural network-based structure is employed, and a novel loss function is formulated by simultaneously considering the monotonicity and curvature of the constructed health index and its variability at the failure time. A tailored adaptive moment estimation algorithm (Adam) is proposed for model parameter estimation. The effectiveness of the proposed method is demonstrated and compared through a case study using the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) data set.

Modeling and analysis of profiles, especially high-dimensional nonlinear profiles, is an important and challenging topic in statistical process control. Conventional mixed-effects models have several limitations in solving the multichannel profile detection problems for in-line Raman spectroscopy, such as the inability to separate defective information from random effects, computational inefficiency, and inability to handle high-dimensional extracted coefficients. In this paper, a new wavelet-based penalized mixed-effects decomposition (PMD) method is proposed to solve the multichannel profile detection problem in Raman spectroscopy. The proposed PMD exploits a regularized high-dimensional regression with linear constraints to decompose the profiles into four parts: fixed effects, normal effects, defective effects, and signal-dependent noise. An optimization algorithm based on the accelerated proximal gradient (APG) is developed to do parameter estimation efficiently for the proposed model. Finally, the separated fixed effects coefficients, normal effects coefficients, and defective effects coefficients can be used to extract the quality features of fabrication consistency, with in sample uniformity, and defect information, respectively. Using a surrogated data analysis and a case study, we evaluated the performance of the proposed PMD method and demonstrated a better detection power with less computational time.

Process monitoring and quality control is essential for continuous manufacturing processes of carbon nano- tube (CNT) thin sheets or buckypaper. Raman spectroscopy is an attractive inline quality characterization and quantification tool for nanomanufacturing because of its nondestructive nature, fast data acquisition speed, and ability to provide detailed material information. However, there is signal-dependent noise buried in the Raman spectra, which reduces the signal-to-noise (S/N) ratio and affects the accuracy, efficiency, and sensitivity for Raman spectrum-based quality control approaches. In this paper, a signal analysis model with signal-dependent noise for Raman spectroscopy is developed and validated based on experimental data. The wavelet shrinkage method is used for denoising and improving the S/N ratio of raw Raman spectra. Based on the validated signal-noise relationship, a novel generalized wavelet shrinkage approach is introduced to remove noise in all wavelet coefficients by applying individual adaptive wavelet thresholds. The effectiveness of this method is demonstrated using both simulation and experimental case studies of inline Raman monitoring of continuous buckypaper manufacturing. The proposed method allows for a significant reduction of Raman data acquisition time without much loss of S/N ratio, which inherently enables Raman spectroscopy for inline monitoring and control for continuous nanomanufacturing processes.