Search

Thermal error modeling (TEM) stands as a pivotal factor in maintaining the machining accuracy of electric spindles. Deep learning (DL) techniques have shown promising potential in this area, however, they face formidable challenges including domain shift issue between training and testing datasets, as well as scarce samples due to the dynamic nature of variable working conditions. Deep transfer learning (DTL) emerges as an encouraging tool, leveraging knowledge obtained from the source domain to elevate learning outcomes within the target domain, notably addressing complicated cross-domain modeling predicaments. Nonetheless, prevailing research largely focuses on enhancing the prediction accuracy within the specific working condition, inadvertently sidestepping the problems of domain shift and scarce sample modeling. To address these issues, a novel weakly supervised adversarial network (WSAN) is proposed for cross-domain TEM with scarce samples. Firstly, the multi-scale convolutional neural network (MSCNN) is constructed to adeptly extract predictive information from multi-sensor data, effectively capturing the time-series patterns embedded within. Furthermore, the adversarial training technique is adopted to address distribution discrepancies that often manifest between distinct domains. To overcome the challenge of scarce labeled samples in the target domain, a weakly supervised learning strategy is employed to deftly control the gradient within the target domain, significantly enhancing the efficacy of positive transfer. The effectiveness and superiority of the proposed method are assessed through extensive experiments using datasets from variable working conditions. The experimental results demonstrate that the proposed method achieves satisfactory performance and outperforms state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

Reducing the number of lower-level function evaluations (LFEs) is crucial to designing a bilevel evolutionary algorithm (BLEA) for bilevel optimization. Direct neighbor solution transfer (DST) is regularly used in BLEA because it can improve the lower-level search efficiency and reduce the number of LFEs. But this method still faces the challenge of misleading searches to make the algorithm converge to the wrong optimum. This paper proposes a bilevel improved multi-operator differential evolution algorithm (BL-IMODE) for bilevel optimization, which uses the knowledge transfer and adaptive coordinate systems approach. Specifically, this paper proposes a lower-level search knowledge transfer (LLS-KT) and a lower-level (LL) optimal solution distribution information transfer (LLO-DIT) mechanism to implement knowledge transfer. In LLS-KT, transferring LL solutions to the LL optimizer is similar to DST, but the LLS-KT transfers multiple solutions and thus has less possibility of misleading searches than DST. Moreover, the LLO-DIT transfers the LL optimal solution distribution information to the LL optimizer for generating promising offspring solutions. Since this method does not directly transfer LL solutions, it avoids misleading searches. Additionally, the adaptive coordinate systems (ACS) approach is used in two-level optimizers to suit different fitness landscapes and further reduce the number of LFEs. In ACS, three differential evolution mutation operators are produced in two coordinate systems. An adaptive strategy is used to select the suitable differential evolution mutation operator to generate promising offspring solutions in the Eigen coordinate or original coordinate system. Experimental studies on three test suites benchmark problems for comparison with seven state-of-the-art algorithms are conducted. The results show BL-IMODE has a significant advantage over the comparative algorithms. Finally, BL-IMODE is applied to tackle the inverse optimal control problem. • We propose a bilevel improved multi-operator differential evolution algorithm. • We propose a knowledge transfer approach to avoid misleading search. • We propose an adaptive coordinate systems approach to improve search efficiency. • We outperform all existing approaches in most test instances. [ABSTRACT FROM AUTHOR]

First-article inspection of flexographic printing labels (FPLs) is significant before the mass production of FPLs, which is manually implemented by quality check (QC) workers in real industries. In this paper, an adaptive coarse-to-fine framework is proposed for automatic first-article inspection of FPLs based on design drawings, consisting of coarse matching, fine matching, and defect detection. Specifically, a deep learning method integrating SuperPoint and SuperGlue is used for coarse matching between the design drawings and the corresponding FPL image to reveal their inherent characteristics. To deal with local deviations of the characters in the FPL image, a character-wise normalized cross-correlation (NCC) method is proposed to finely align each character in the design drawings with that in the FPL image, which utilizes different searching windows for different characters. Next, a constrained clustering method is proposed to accurately extract the contents of interest (COIs) in the FPL image. Finally, an edge distance measure with an adaptive threshold scheme is proposed to evaluate the COIs for inspection. Comparison experiments show that the proposed framework achieves better inspection performance for first-article inspection of FPLs than other fabric inspection methods. Specifically, it achieves the inspection performance of 1.3% missing rate, 2.0% false rate, and 0.951 Dice at an acceptable inspection speed of 2.735 s/pcs, which demonstrates that it is applicable to facilitate the QC task in real industries. [ABSTRACT FROM AUTHOR]