Advances in Manufacturing最新文献

In lightweight automotive vehicles, the application of self-piercing rivet (SPR) joints is becoming increasingly widespread. Considering the importance of automotive service performance, the fatigue performance of SPR joints has received considerable attention. Therefore, this study proposes a data-driven approach to predict the fatigue life and failure modes of SPR joints. The dataset comprises three specimen types: cross-tensile, cross-peel, and tensile-shear. To ensure data consistency, a finite element analysis was employed to convert the external loads of the different specimens. Feature selection was implemented using various machine-learning algorithms to determine the model input. The Gaussian process regression algorithm was used to predict fatigue life, and its performance was compared with different kernel functions commonly used in the field. The results revealed that the Matern kernel exhibited an exceptional predictive capability for fatigue life. Among the data points, 95.9% fell within the 3-fold error band, and the remaining 4.1% exceeded the 3-fold error band owing to inherent dispersion in the fatigue data. To predict the failure location, various tree and artificial neural network (ANN) models were compared. The findings indicated that the ANN models slightly outperformed the tree models. The ANN model accurately predicts the failure of joints with varying dimensions and materials. However, minor deviations were observed for the joints with the same sheet. Overall, this data-driven approach provided a reliable predictive model for estimating the fatigue life and failure location of SPR joints.

This paper presents a new machine learning-based calibration framework for strength simulation models of self-piercing riveted (SPR) joints. Strength simulations were conducted through the integrated modeling of SPR joints from process to performance, while physical quasi-static tensile tests were performed on combinations of DP600 high-strength steel and 5754 aluminum alloy sheets under lap-shear loading conditions. A sensitivity study of the critical simulation parameters (e.g., friction coefficient and scaling factor) was conducted using the controlled variables method and Sobol sensitivity analysis for feature selection. Subsequently, machine-learning-based surrogate models were used to train and accurately represent the mapping between the detailed joint profile and its load-displacement curve. Calibration of the simulation model is defined as a dual-objective optimization task to minimize errors in key load displacement features between simulations and experiments. A multi-objective genetic algorithm (MOGA) was chosen for optimization. The three combinations of SPR joints illustrated the effectiveness of the proposed framework, and good agreement was achieved between the calibrated models and experiments.

Class imbalance is a common characteristic of industrial data that adversely affects industrial data mining because it leads to the biased training of machine learning models. To address this issue, the augmentation of samples in minority classes based on generative adversarial networks (GANs) has been demonstrated as an effective approach. This study proposes a novel GAN-based minority class augmentation approach named classifier-aided minority augmentation generative adversarial network (CMAGAN). In the CMAGAN framework, an outlier elimination strategy is first applied to each class to minimize the negative impacts of outliers. Subsequently, a newly designed boundary-strengthening learning GAN (BSLGAN) is employed to generate additional samples for minority classes. By incorporating a supplementary classifier and innovative training mechanisms, the BSLGAN focuses on learning the distribution of samples near classification boundaries. Consequently, it can fully capture the characteristics of the target class and generate highly realistic samples with clear boundaries. Finally, the new samples are filtered based on the Mahalanobis distance to ensure that they are within the desired distribution. To evaluate the effectiveness of the proposed approach, CMAGAN was used to solve the class imbalance problem in eight real-world fault-prediction applications. The performance of CMAGAN was compared with that of seven other algorithms, including state-of-the-art GAN-based methods, and the results indicated that CMAGAN could provide higher-quality augmented results.

This research presents an innovative approach to accurately predict the nugget diameter in resistance spot welding (RSW) by leveraging machine learning and transfer learning methods. Initially, low-fidelity (LF) data were obtained through finite element numerical simulation and design of experiments (DOEs) to train the LF machine learning model. Subsequently, high-fidelity (HF) data were collected from RSW process experiments and used to fine-tune the LF model by transfer learning techniques. The accuracy and generalization performance of the models were thoroughly validated. The results demonstrated that combining different fidelity datasets and employing transfer learning could significantly improve the prediction accuracy while minimize the costs associated with experimental trials, and provide an effective and valuable method for predicting critical process parameters in RSW.

The stress-life curve (S–N) and low-cycle strain-life curve (E–N) are the two primary representations used to characterize the fatigue behavior of a material. These material fatigue curves are essential for structural fatigue analysis. However, conducting material fatigue tests is expensive and time-intensive. To address the challenge of data limitations on ferrous metal materials, we propose a novel method that utilizes the Random Forest Algorithm and transfer learning to predict the S–N and E–N curves of ferrous materials. In addition, a data-augmentation framework is introduced using a conditional generative adversarial network (cGAN) to overcome data deficiencies. By incorporating the cGAN-generated data, the accuracy (R²) of the Random Forest Algorithm-trained model is improved by 0.3–0.6. It is proven that the cGAN can significantly enhance the prediction accuracy of the machine-learning model and balance the cost of obtaining fatigue data from the experiment.

The exoskeleton is employed to assist humans in various domains including military missions, rehabilitation, industrial operation, and activities of daily living (ADLs).Walking is a fundamental ADL, and exoskeletons are capable of reducing the activation and metabolism of lower extremity muscles through active assistance during walking. To improve the performance of active assistance strategy, this article proposes a framework using an active hip exoskeleton. Subsequently, it correlates to an already established Bayesian-based human gait recognition algorithm, with a particular focus on linear and circular walking within industrial and ADL contexts. In theorizing this strategy for exoskeletons, this study further reveals, in part, the activation characteristics of human hip muscles for the instruction and regulation of active assistance duration and onset timing. This proposed active assistance strategy provides new insights for enhancing the performance of assistive robots and facilitating human robot interaction within the context of ADLs.

Roping is a severe band-like surface defect that occurs in deformed aluminum alloy sheets. Accurate roping prediction and rating are essential for industrial applications. Recently, the authors introduced an artificial neural network (ANN) model to efficiently forecast roping behavior across the thickness of large regions with texture gradients. In this study, the previously proposed ANN model for roping prediction is briefly reviewed, and a few-shot learning (FSL)-based method is developed for roping grading with limited samples. To consider the directionality of the roping patterns, the roping dataset constructed from experimental observations is transformed into the frequency domain for more compact characterization. A transfer-based FSL method is further presented for grade roping with manifold mixup regularization and the Sinkhorn mapping algorithm. A new component-focused representation is also implemented for data-processing, exploiting the close correlation between roping and power distribution in the frequency domain. The ultimate FSL method achieved an optimal accuracy of 95.65% in roping classification with only five training samples per class, outperforming four typical FSL methods. This FSL approach can be applied to grade the roping morphologies predicted by the ANN model. Consequently, the combination of prediction and grading using deep learning provides a new paradigm for roping analysis and control.

Wax pattern fabrication in the investment casting of hollow turbine blades directly determines the dimension accuracy of subsequent casting, and therefore significantly affects the quality of final product. In this work, we develop a machine learning-based multi-objective optimization framework for improving dimension accuracy of wax pattern by optimizing its process parameters. We consider two optimization objectives on the dimension of wax pattern, i.e., the surface warpage and core offset. An active learning of Bayesian optimization is employed in data sampling to determine process parameters, and a validated numerical model of injection molding is used to compute objective results of dimension under varied process parameters. The collected dataset is then leveraged to train different machine learning models, and it turns out that the Gaussian process regression model performs best in prediction accuracy, which is then used as the surrogate model in the optimization framework. A genetic algorithm is employed to produce a non-dominated Pareto front using the surrogate model in searching, followed by an entropy weight method to select the most optimal solution from the Pareto front. The optimized set of process parameters is then compared to empirical parameters obtained from previous trial-and-error experiments, and it turns out that the maximum and average warpage results of the optimized solution decrease 26.0% and 20.2%, and the maximum and average errors of wall thickness compared to standard part decrease from 0.22 mm and 0.051 7 mm using empirical parameters to 0.10 mm and 0.035 6 mm using optimized parameters, respectively. This framework is demonstrated capable of addressing the challenge of dimension control arising in the wax pattern production, and it can be reliably deployed in varied types of turbine blades to significantly reduce the manufacturing cost of turbine blades.