Abstracts:Convolutional neural networks (CNNs) have demonstrated a great potential in fault detection for a wide type of multilevel inverters. Despite the remarkable performance of CNNs, their interpretability remains a challenge. This is due to networks, which have complicated black boxes behaviors. Consequently, they present a substantial challenge for widespread adoption of different models in practical applications. Moreover, relying solely on accuracy is insufficient, especially in critical applications where maintaining trust and robustness is vital for protecting a system against potential damage. Therefore, this study implements a visual explanation method called gradient weighted class activation map (Grad-CAM) for fault detection of multilevel inverter. The Grad-CAM method can identify the model’s important features and interpret the detection of fault types. The proposed method was validated by both simulation and experimental results for three-level neutral-point clamped inverters, demonstrating that a reliable CNN achieved high classification accuracy and effectively identified fault types.

Abstracts:In-situ metrology of the pre-etching geometry of printed patterns is critical for roll-to-roll microcontact printing (R2R $\mu$CP) quality control. In the previous works, condensation figures (CFs) were successfully used to visualize the nanometer-thick printed layers in R2R $\mu$CP. However, applying CFs for real-time monitoring and controlling the R2R $\mu$CP remains highly challenging. This is primarily due to the difficulty in extracting printed features from CFs, as pattern segmentation is often time-consuming and inaccurate due to the complexity and instability of the droplet size and intensity distribution. To solve this problem, we leverage the outstanding efficiency and accuracy of the you-only-look-once computer vision model to measure printed patterns from droplet-only CFs by segmenting them into printed and nonprinted areas. To address the difficulty of collecting CFs of real R2R $\mu$CP with annotated labels for segmentation tasks, we develop a simulation pipeline to generate, augment, and label synthetic CFs for training deep image segmentation models. Experimental results on real R2R $\mu$CP processes show that the proposed method achieves a 94.4% accuracy in line-width estimation, based on comparisons with postetching measurements, and outperforms other state-of-the-art deep learning and classic segmentation techniques. Finally, the method was integrated in an R2R $\mu$CP system, enabling fully automated in-situ metrology at up to 15 CFs per second with 1080p resolution, providing real-time feedback for closed-loop quality control.

Abstracts:Wind farm layout and yaw control both play key roles in maximizing energy yield. However, these two factors are often optimized separately, which can lead to less effective designs. This article proposes a new biobjective optimization method for wind farm layout that includes yaw control. The two main goals are to maximize annual energy production (AEP) and minimize land use. To efficiently solve the lower-level yaw control problem, a hierarchical search method (HSM) is proposed. This method can quickly provide nearly optimal yaw settings for any specific layout. In the upper-level layout optimization, the epsilon-constraint method is used alongside sequential quadratic programming (SQP). The proposed method is tested with simulations on the FLORIS platform, using NREL 5 MW reference turbines. Compared to traditional layout optimization method, the integrated layout and yaw optimization achieves the highest AEP across all tested land area limits. Compared to the gradient-based method, our proposed HSM can avoid getting trapped in local optima and has a much faster computation time in the power optimization problem.

Abstracts:Faults in complex industrial equipment typically exhibit continuous evolutionary characteristics. However, existing diagnostic methods are generally limited to discrete level fault classification, making it challenging to cover the full spectrum of fault severities and thereby restricting the generalization capability of diagnostic models. This article proposes a novel approach that models the fault evolution process to generate fault data corresponding to unseen level faults, enabling the diagnosis of these unseen level faults. Specifically, within the framework of a temporal variational autoencoder, we employ neural controlled differential equations combined with gated recurrent units as the temporal encoder to capture the posterior distribution of the data. Simultaneously, the prior distribution in the latent space is modeled using the Koopman operator, allowing for linear modeling of the latent dynamics associated with fault evolution and uncovering the underlying patterns of system degradation. By sampling latent representations corresponding to unseen level faults, we generate synthetic fault data to train a model capable of diagnosing unseen level faults. Experimental results on a ground source heat pump system demonstrate that the proposed method can generate high-quality data for unseen level faults and exhibits strong performance in diagnosing unseen level faults.

Abstracts:Social manufacturing integrates social resources with manufacturing, demanding rapid decision-making by processing specialized and complex data to adapt market changes. Leveraging multisource information as external knowledge bases, implementing Retrieval-Augmented Generation (RAG) to Large Language Models (LLMs) holds great potential to enhance the manufacturing efficiency. However, existing RAG methods often return excessive unstructured information, limiting LLMs’ ability to reason and solve complex question-and-answer (QA) tasks. Drawing inspiration from structured human learning, we propose a novel Curriculum Enhanced Retrieval-Augmented Generation (CE-RAG) framework. CE-RAG encompasses three stages: The syllabus setting phase for organizing subjects and study sequences for the knowledge base, the knowledge infilling phase for generating curriculum chain prompts for LLMs, and the curriculum review phase for memorizing the curriculum chains learned by LLMs to improve iterative retrieval. Experiments on public QA reasoning datasets and social manufacturing cases show CE-RAG can effectively improve the reasoning performance of LLMs in complex social manufacturing QA tasks.

Abstracts:Accurate battery capacity estimation is key to alleviating consumer concerns about battery performance and reliability of electric vehicles (EVs). However, practical data limitations imposed by stringent privacy regulations and labeled data shortages hamper the development of generalizable capacity estimation models that remain robust to real-world data distribution shifts. While self-supervised learning can leverage unlabeled data, existing techniques are not particularly designed to learn effectively from challenging field data—let alone from privacy-friendly data, which are often less feature-rich and noisier. In this work, we propose a first-of-its-kind capacity estimation model based on self-supervised pretraining, developed on a large-scale dataset of privacy-friendly charging data snippets from real-world EV operations. Our pre-training framework, snippet similarity-weighted masked input reconstruction, is designed to learn rich, generalizable representations even from less feature-rich and fragmented privacy-friendly data. Our key innovation lies in harnessing contrastive learning to first capture high-level similarities among fragmented snippets that otherwise lack meaningful context. With our snippet-wise contrastive learning and subsequent similarity-weighted masked reconstruction, we are able to learn rich representations of both granular charging patterns within individual snippets and high-level associative relationships across different snippets. Bolstered by this rich representation learning, our model consistently outperforms state-of-the-art baselines, achieving 31.9% lower test error than the best-performing benchmark, even under challenging domain-shifted settings affected by both manufacturer and age-induced distribution shifts.

Abstracts:The rapid development of quantum computing poses a significant threat to the security of the Industrial Internet of Things (IIoT), rendering traditional cryptographic systems inadequate for ensuring long-term security. As a fundamental technology for establishing trust in IIoT networks, digital signatures are essential for secure device authentication, data integrity verification, and the protection of communication channels. However, implementing efficient digital signature schemes in resource-constrained embedded devices presents significant challenges, particularly when faced with the demands of postquantum cryptography (PQC). We propose a scheme for resource-intensive GPU optimization for HAETAE throughput tuning (RIGHT). Specifically, we employ coarse-grained parallelism and kernel fusion to maximize the parallel processing capabilities of GPUs. Furthermore, we present a hierarchical data locality optimization method for NTT/INTT and FFT, and optimize the SHAKE256 using CUDA PTX instructions, further enhancing throughput to meet the high concurrency and computational performance requirements of IIoT applications. On the Jetson Xavier, the signing performance of HAETAE is 1.25 x, 1.09 x, and 1.31 x that of the Intel i7-10700 K CPU with AVX2, while verification performance is about 4 times faster, demonstrating low power consumption and high efficiency, making it suitable for edge computing. Specifically, on the NVIDIA RTX 4090, the signature throughput reaches 155 324 ops/s, and the verification throughput reaches 4 276 075 ops/s, showcasing significant parallel computing capabilities, which is ideal for large-scale digital signature verification tasks in IIoT.

Abstracts:Understanding human motion is fundamental to the development of intelligent systems capable of seamless interaction with people. Trajectory prediction is a critical component in domains, such as intelligent transportation, surveillance, and human–robot collaboration. However, accurately forecasting human movement remains a significant challenge due to its inherently uncertain and multimodal nature. In this work, we propose a deep neural network that models agent dynamics and predicts future trajectories by representing them as a probabilistic multimodal distribution. To effectively capture the stochasticity of human behavior, our method employs a score-based diffusion model that learns to generate realistic trajectory samples by denoising latent representations. In addition, we introduce a novel social attention mechanism designed to model complex interagent interactions, further improving predictive performance. We validate our approach in both pedestrian and marine vessel trajectory datasets, demonstrating its superior ability to capture social dynamics and forecast diverse plausible future outcomes. Extensive experiments and ablation studies confirm the robustness, generalizability, and accuracy of our framework in varied real-world environments.

Abstracts:In industrial inspection, severe class imbalance poses a major challenge for training accurate detection models. Generative methods offer a potential solution by augmenting minority-class images; however, most existing approaches exhibit poor generalization under extremely low-shot conditions and often produce semantically implausible results. To address these limitations, a human-in-the-loop generative framework, HuIV-GAN, is proposed. It integrates a variational autoencoder with a generative adversarial network to synthesize images under conditions of severe data scarcity. HuIV-GAN incorporates real-time human feedback to refine generation iteratively: experts assess intermediate outputs, enabling a closed-loop generate–filter–retrain process that ensures both structural realism and semantic accuracy. The proposed method is validated on a typical detection task to solve the scarcity of solder joint fracture data in ball grid array (BGA) packaging. The red-dye BGA dataset has a highly imbalanced label distribution, with the smallest minority class constituting less than 0.1%. Our method significantly improved minority class detection under extreme imbalance, whereas detection models trained on the original dataset failed to detect this class entirely. Furthermore, the feedback loop reduced training instability and accelerated convergence by over 200 iterations.

Abstracts:The recent amendment to the IEEE 1547 standard enables smart inverters to support grid functions like Volt-VAR/Watt and frequency Watt control, which necessitates communication between smart inverters and a microgrid central communication controller. The IEEE 1547 standard recommends this communication through Sunspec Modbus protocol. However, this standardization of communication protocol increases vulnerability to cyber-attacks. Sunspec Modbus, lacking built-in security measures, exacerbates this risk by further expanding the potential attack surface within microgrids. To determine the cyber-attack path in smart inverters, a comprehensive cyber-attack model for exploiting the cybersecurity limitations in Sunspec Modbus protocol is developed. A real-time high-fidelity cyber-physical energy system (CPES) testbed is developed to assess the impact of cyber-attacks on smart inverter control. Finally, to mitigate these cyber-attacks, a novel lightweight cybersecurity scheme based on Blake-2 s and AES-256 algorithm for the Sunspec Modbus protocol is proposed. The effectiveness of the proposed security scheme to mitigate the cyber-attacks is evaluated on the developed CPES testbed. Due to the lightweight feature of both algorithms, the proposed scheme is found to be suitable for low computation platforms, such as smart inverters.