Wei-Jen Wang

Defect classification for thin-film transistor liquid crystal displays (TFT-LCD) poses significant challenges due to the fine-grained nature of microscopic defects and the limited availability of labeled data. This problem is particularly demanding in the context of few-shot class-incremental learning (FSCIL). FSCIL requires learning new defect classes from limited data while distinguishing between highly similar classes. At the same time, it must preserve knowledge of previously learned classes. We introduce Fine-grained Few-shot Class Incremental Learning with Destruction-construction Learning (FFCIL) to address these challenges. This approach employs a multistage training strategy that begins with destruction-construction learning. This stage captures intricate and fine-grained features during pretraining and is combined with manifold-mixup as a placeholder mechanism. Together, these components enable the seamless integration of new defect classes with minimal samples. Furthermore, a triplet loss is applied during pseudo-incremental learning to improve the model’s ability to differentiate between base and pseudo-classes. This approach enhances the fine-grained classification performance in a few-shot class incremental learning scenarios. Experimental results on an industry-based TFT-LCD dataset demonstrate a notable 16.36% improvement in overall accuracy. In addition, FFCIL achieves the lowest performance dropping rate among all compared methods, indicating superior stability. Tests on public datasets further reveal an improvement of up to 39.58% in overall accuracy, underscoring the potential applicability of FFCIL to broader domains. Code is available at https://github.com/sallylinlab/ffcil

In industrial production, automated inspection methods for circular knitting machines often encounter several challenges. First, the rapid movement of fabrics on these machines makes it difficult for existing fabric defect detection methods to effectively capture and process the motion. Next, due to practical constraints aimed at maintaining high yield rates, collecting sufficient abnormal fabric samples for model training is costly and limited. Furthermore, circular knitting machines typically operate under varying illumination conditions, further complicating the task of accurate fabric defect detection. Additionally, these methods usually fail to identify the cutline patterns that are integral to the design of the fabric and mistake cutlines for v-line defects. Therefore, existing fabric defect detection methods often struggle to balance real-time processing, few-shot learning, and high accuracy under various illumination conditions To address the aforementioned challenges, we adopt a few-shot learning approach and propose a novel real-time fabric defect detection method for circular knitting machines, aiming to achieve high accuracy even under varying illumination conditions. The proposed mechanism consists of two components, the LBUnet and the false alarm filter for cutlines. First, to tackle the challenges of real-time detection, limited training data, and varying illumination conditions, we develop a lightweight semantic segmentation model, LBUnet, which leverages local binary (LB) convolution to effectively handle variable lighting conditions. Next, to address the specific challenge of detecting V-line defects, we propose a false-alarm filtering method that ensures accurate defect identification by utilizing time-series data composed of consecutive segmentation maps generated by LBUnet. Extensive experiments demonstrate that the proposed method delivers both high defect detection accuracy and real-time processing performance for fast-moving fabrics on circular knitting machines under diverse lighting conditions. Specifically, using only LBUnet, our approach achieved an average Mean Intersection over Union (mIoU) of 86.24% with an average processing time of just 4 milliseconds per image. When the false-alarm filtering component was incorporated, the system achieved 100% accuracy in detecting cutlines.

The web shell attacks (WebShells) have long been a source of persistent annoyance for administrators. They have become a major security concern in cloud computing environments since the scalability and distributed nature of the cloud services could intensify the potential risks and impacts of such attacks. In response, researchers have proposed numerous strategies to shield assets from WebShell intrusions. Consequently, this study proposes a method that utilizes the BPE (Byte Pair Encoding) for tokenization, the CodeBERT for extracting the word embedding vector of a given source code piece, and a deep model (GRU or Bidirectional GRU) for determining whether the code contains a WebShell. This architecture is designed to detect the presence of WebShells in PHP code through analysis of the source code. Our experimental results indicate that, the proposed method with GRU achieves the best performance, with an accuracy of 99.72%, a precision of 99.36%, and an F1-score of 99.36%. Furthermore, it outperforms the methods proposed in the prior related studies as presented in the paper.

In batch production systems, detecting low-yield machines is essential for minimizing the production of defective pieces, which is a complex problem that currently requires multiple experts, considerable capital, or a combination of both to overcome. To solve this problem, we proposed a cost-efficient and straightforward method that involves using maximum likelihood estimation and bootstrap confidence intervals to estimate per-machine yield; this method enables identification of low-yield machines and generation of a list of these machines. Manufacturing engineers can use the list to perform necessary verification and maintenance processes. Before implementing this method, a manufacturer with 50–500 machines should build a dataset containing approximately 6–20 times as many batches as there are production machines. When this condition is met, the proposed method can be used effectively to detect up to five low-yield machines.

Defects in multistage manufacturing processes (MMPs) decrease profitability and product quality. Therefore, MMP parameter optimization within a range is essential to prevent defects, achieve dynamic accuracy, and accommodate manufacturing tolerances. However, existing studies only focused on optimization in a single manufacturing stage of MMP, such as the weaving stage in fabric manufacturing. Furthermore, existing methods optimize for a single value rather than a range. Thus, we propose a novel approach called multistage parameter optimization for rule generation (MPORG) to prevent the occurrence of defects in MMPs. In the proposed approach, key parameters are identified and optimized to a range for each defect type. Subsequently, the optimized parameters for each defect type are merged. Our approach is novel because it optimizes parameters to a range rather than a single value, allowing engineers to select a value in this range according to their experience. It also provides results that are specific to a product type. Our approach outperformed the classification and regression tree (CART) algorithm and multiresponse CART method in experiments on an empirical fabric manufacturing dataset that we gathered. The experimental results demonstrated that the MPORG approach can prevent the occurrence of single-type or multiple-type defects by approximately 89%.