8 July 2020 in IEEE Transactions on Industrial Informatics
The automatic defects detection for solar cell electroluminescence (EL) images is a challenging task, due to the similarity of defect features and complex background features. To address this problem, a novel Complementary Attention Network (CAN) is designed by connecting the novel channel-wise attention subnetwork with spatial attention subnetwork sequentially, which adaptively suppresses the background noise features and highlights the defect features simultaneously by employing the complementary advantage of the channel features and spatial position features. In CAN, the novel channel-wise attention subnetwork applies convolution operation to integrate the concatenated and discriminative output features extracted by Global Average Pooling (GAP) layer and Global Max Pooling (GMP) layer, which can make fully use of these informative features. Furthermore, a Region Proposal Attention Network (RPAN) is proposed by embedding CAN into Region Proposal Network (RPN) in Faster R-CNN to extract more refined defective region proposals, which is used to construct a novel end-to-end Faster RPAN-CNN framework for detecting defects in raw EL image. Finally, some experimental results on a large-scale EL dataset including 3629 images, 2129 of which are defective, show that the proposed method performs much better than other methods in terms of defects classification and detection results in raw solar cell EL images.
20 March 2019 in IEEE Transactions on Instrumentation and Measurement
2 February 2021 in IEEE Transactions on Instrumentation and Measurement
The small hot spot defect detection for photovoltaic (PV) farms is a challenging problem due to the feature vanishing as the network deepens. To solve this challenging problem, a novel residual channelwise attention gate network (RCAG-Net) is proposed by employing a novel RCAG module to achieve multiscale feature fusion, complex background suppression, and defect feature highlighting. In RCAG-Net, the novel RCAG module first realizes feature fusion by adding the features of different scale layers. Next, global average pooling (GAP) and multilayer perceptron (MLP) are used to dimension reduction and refinement of the fused features, then yielding an attention map for channelwise feature reweighting by gate mechanism, which employs selective transmission of the convolution neural network (CNN)-extracted features to achieve informative feature filtering. Moreover, residual connection from the fused features to the final output facilitates the insertion of the new RCAG into some classical pretrained models, without breaking its initial behavior. Finally, the proposed approach is validated through a real defect detection system, and the experimental result clearly verifies its effectiveness for small hot spot detection of PV farms.
6 December 2023 in Measurement
17 April 2025 in IEEE Sensors Journal
The prediction of the Remaining Useful Life (RUL) of stochastically degrading devices is of critical importance for Prognostics and Health Management (PHM). However, most of the existing approaches overlook the inherent coupling of trends and fluctuations in the degradation process. Meanwhile, these methods are struggling to effectively model the random fluctuations component in the degradation process and may lead to inaccurate prediction. To address these limitations, we propose a modified RUL prediction framework WTSAGNN. The framework extends the spatiotemporal dependencies of multi-sensor signals into the frequency domain. By leveraging the proposed adaptive orthogonal wavelet convolution to more effectively decouple the trend and fluctuation components of the degradation process. Furthermore, we introduce a stochastic differential network (SDN) to replace the traditional multilayer perceptron (MLP). The proposed method enhances the robustness of neural networks against stochastic variations in degradation processes through explicit stochastic modeling. Experimental results demonstrate that WTSAGNN outperforms existing state-of-the-art(SOTA) methods on benchmark datasets, achieving up to a 9.1% improvement in prediction accuracy with only a limited increase in computational overhead. In addition, the proposed modules are plug-and-play, allowing seamless integration with existing data-driven prediction frameworks.
1 January 2026 in IEEE Transactions on Automation Science and Engineering