Search

Background: The recent advances in the manufacturing industry, mainly fueled by the fourth industrial revolution, and the shortage of skilled welders are boosting research towards intelligent automatic welding. In this context, vision sensing and machine learning have been widely applied to the development of welding status recognition (WSR) systems. However, researchers have mainly focused on developing accurate models putting aside the analysis on computational performance, which is a relevant aspect of the development of real-time WSR systems. Objective: This work aims at devising improved modeling paradigms, considering optimization towards both accuracy and computational performance, for real-time WSR in gas tungsten arc welding. Methods: Shallow learning paradigms are compared with a benchmark deep learning model regarding classification as well as computational performance. For that purpose, metrics that account for the time a model takes to make a prediction and the hard disk space occupied by the model are proposed. The dataset used in this paper comprehends molten pool images acquired for defective and non-defective welding conditions. A dimensionality reduction technique was used to extract features from the data efficiently. For the model creation, a randomized hyperparameter search strategy was adopted. Results: The modeling workflow adopted resulted in an efficient and lightweight model suitable for being deployed as a diagnostic tool. Results have shown that shallow learning jointly with principal component analysis improved both the accuracy and hardware consumption. Specifically, compared to the benchmark deep learning model, the developed classifier presented size 84.82% smaller, and related accuracy improved by 17.32%. Conclusion: The proposed approach has shown to be an efficient way of achieving a compromise between classification and computational performance when developing a predictive model to serve as an online tool for vision-sensing-based WSR. [ABSTRACT FROM AUTHOR]

Background: Due to the acoustoelastic effect, ultrasonic guided waves have been used to estimate mechanical stress in a cheap and nondestructive fashion. Machine learning has been applied to map complex waveforms to stress estimates, though important aspects to construct and deploy real-time monitoring systems, such as accuracy and hardware consumption, to date, have not been concomitantly explored. Objective: The goal of the present paper is to propose a data modeling methodology that optimizes accuracy and computational implementation, towards devising the best practices for real-time ultrasonic-based stress estimation. Methods: We evaluate shallow and deep learning models with dimensionality reduction, which are compared to the most recent work showing overall better results for the approach presented herein both in terms of accuracy and hardware resources consumption. We generated a dataset to evaluate the models on a test rig with a plate measuring (i) ultrasonic guided waves excited by broadband signals and (ii) different stress conditions. The models are created and tested using a Monte-Carlo hold-out procedure with a repeated k-fold randomized hyperparameter search to evaluate their robustness in different stress conditions. Results: We dramatically improve the current state of the art by proposing and evaluating a more efficient model construction procedure. Results show that by using shallow models and principal component analysis, we were able to improve the model performance in terms of predictive power and inference hardware consumption. Specifically, we obtained an accuracy improvement of about 10% and hardware consumption used for inference smaller in three orders of magnitude when compared to the state-of-the-art reported with deep neural network models. Conclusions: Results herein depicted impact the way practitioners may proceed for building efficient embedded predictive models for stress estimation using guided waves, enabling more precise, decentralized, and real-time nondestructive monitoring. [ABSTRACT FROM AUTHOR]

Background: A variety of imaging methods are available to obtain kinematic data at an interface, with a widely varying range of spatial and temporal resolution. These methods require a trade-off between imaging rate and resolution. Objective: A deep learning framework trained on synchronous profilometry data acquired using two imaging modalities at two different spatial resolutions to enhance spatial resolution while maintaining temporal resolution is desired. Methods: Fizeau interferometry (FIF) and frustrated total internal reflection (FTIR) are used to overcome the resolution-rate trade-off via a deep learning framework. The FTIR imaging data are recorded at high resolution, while the FIF imaging data are recorded with a lesser resolved, larger field of view. We apply a deep learning framework using a multi-layer convolutional neural network to enhance the FIF image resolution. Results: With the deep learning framework, we achieve the high spatial resolution of measurements obtained by FTIR imaging in all three dimensions from the lower resolution FIF data. A high-order overset technique ultimately yields full up-scaled images from the network outputs without losing precision. The accuracy of the super-resolved image is evaluated using test data. Conclusions: This hybrid framework, called HOTNNET, is implemented in its entirety on high-speed imaging profilometry data acquired in the study of droplet impacts on a smooth, solid surface, and is used to recover full, high-resolution images at high rates by unwrapping the phase of the interferometry. This framework can be readily adapted to other paired datasets by retraining the network on the novel data. [ABSTRACT FROM AUTHOR]

Background: In digital image correlation (DIC), high-quality speckle patterns and quality assessments have attracted significant attention because successfully performing these assessments could be a key challenge. Previous studies have pointed out that a speckle pattern with a more plentiful gradient generally leads to higher local uniqueness, smaller bias errors and reduced random errors. Most of existing metrics in the literature were based on such qualitative prior knowledge of the speckle-pattern features in DIC measurements. Objective: In this study, a new method using convolution neural networks (CNNs) is proposed to assess speckle patterns without any pre-existing qualitative knowledge while providing better performance than other existing speckle pattern metrics. Methods: Speckle-pattern images of various sizes were fabricated as inputs for training data, and three displacement fields were numerically employed to obtain the average errors of the DIC measurements as outputs. The data sets for training and testing were designed to investigate the learning capability of the CNNs for challenging tasks. The CNNs architecture is explained along with the detailed composition of its layers that were employed in this study. After training, CNNs can learn the underlying relationship between speckle pattern images and target errors, without the need for any pre-existing qualitative or quantitative knowledge about the speckle patterns. Results: The CNN error prediction after training shows a higher correlation coefficient with the target than existing measures, namely, mean intensity gradient (MIG), mean subset fluctuation (MSF), Shannon entropy (SE), and standard deviation of gray intensities within each speckle (SDGIS). The accuracy of the existing measures generally decreased as the image size decreased. However, the correlation coefficient resulting from the CNN prediction not only provides a higher accuracy in all sizes of the images, but the coefficient also was not significantly affected by the image size. Conclusions: This is probably the first attempt to investigate the effectiveness of using a deep learning-based CNN in the quality assessment of speck patterns. The convolution layer in the CNN has local receptive fields and shared weights to extract local features, and these local features are combined by the subsequent layers to detect their global correlations. These features of using CNN are closely related to the local and global assessment of the quality of speckle patterns. Our study builds on the advantages of CNNs, which can be used as an alternative metric to assess speckle-pattern quality. [ABSTRACT FROM AUTHOR]