ABSTRACT. Driven by smart factories and Industry 4.0, the manufacturing sector is placing higher demands on the speed, accuracy, and traceability of anomaly diagnosis. However, existing approaches still rely heavily on manual interpretation and fragmented data, making it difficult to systematically integrate knowledge and automatically generate improvement plans. In this study, we propose an intelligent anomaly diagnosis system, AI-assisted Industry Anomaly Diagnosis and Correction (AIIA-DC) System, that integrates data digitization, statistical analysis, machine learning, and generative AI models to achieve end-to-end automation from data input to actionable recommendations. The system utilizes principal component analysis (PCA) with correlation coefficients (ΔCorr) and multivariate indices to explain the sources of anomalies, and employs a retrieval-augmented generation (RAG) framework to produce traceable and verifiable diagnostic reports. Using handwritten raw data from an electronic carrier tape manufacturing plant, the proposed method outperforms conventional PCA and standalone machine learning models in anomaly localization and interpretation of feature contributions, demonstrating higher accuracy and practical framework for a traditional industry pursuing for digitalization. The results provide process engineers with a tool for rapid diagnosis, reduced misjudgment, and improved decision-making efficiency, while also establishing a technical foundation for data-driven management and knowledge accumulation in smart factories.

ABSTRACT. The livestock industry faces continuous pressure to enhance efficiency while encountering severe labor shortages and an increasing need for precise health management. Traditional methods for monitoring swine growth and behavior rely heavily on manual observation, which is not only labor-intensive but also susceptible to human error. To address this challenge, this paper proposes a Smart Swine Farming framework that integrates Arti-ficial Intelligence (AI) vision and Internet of Things (IoT) technologies. The framework comprises two core functional modules: a non-contact swine body length estimation module and a dynamic swine activity tracking module. In the swine body length estimation stage, this study utilizes YOLOv11-Pose for keypoint detection and integrates the MiDaS depth model to achieve swine body length calculation without physical contact. In the swine activity tracking stage, the system integrates YOLOv11 with the BoT-SORT tracker to continuously monitor swine behavior patterns, such as feeding, drinking, and corner dwelling. The core functionalities have been encapsulated as a RESTful API and visualized through Virtual Reality (VR) and a web interface, enabling swine farmers to intuitively access man-agement data. The expected outcomes are to significantly enhance the effec-tiveness of swine health management, optimize swine farming efficiency, and reduce labor resource investment for livestock producers.

ABSTRACT. Campus energy management faces continuous pressure to enhance sustainability while maintaining public safety standards. Traditional methods for controlling streetlights rely heavily on timer-based schedules or passive sensors, which are not only rigid in adaptability but also inefficient in energy utilization during low-traffic periods. To address this challenge, this paper proposes a Data-Driven Smart Streetlight framework that integrates Artificial Intelligence (AI) vision and Internet of Things (IoT) technologies. The framework comprises a hybrid edge-cloud architecture: at the edge, a Raspberry Pi-based platform uti-lizes YOLO object detection to fuse heterogeneous sensor data (LDR, PIR), en-abling precise situational awareness. Specifically, a multi-modal brightness con-trol algorithm is employed to dynamically adjust illumination based on real-time pedestrian density, incorporating a temporal smoothing mechanism to mitigate abrupt flickering. The core functionalities regarding data transmission are opti-mized using lightweight MQTT and HTTP MJPEG protocols, and visualized through a unified ELK stack interface, enabling administrators to intuitively ac-cess management data. Experimental results demonstrate that the system achieves a low end-to-end latency of approximately 104 ms and significantly re-duces energy consumption by up to 66% in low-traffic scenarios compared to conventional lighting.

ABSTRACT. Colorectal cancer has a high incidence and mortality rate in Taiwan. Clinically, diagnosis is typically performed using colonoscopy and computed tomography (CT). Compared with colonoscopy, which allows direct visualization and biopsy but often causes significant discomfort that may reduce patient compliance, CT colonography is noninvasive, provides high-resolution images rapidly, and enables visualization of intestinal structures to assist in tumor localization. This retrospective study collected data from 237 patients. CT images were annotated for colon regions and tumor locations. Segmentation models based on nnU-Net and YOLO were developed to delineate the colon and colorectal tumors. To enhance tumor segmentation accuracy, the initially segmented colon region was used as a mask to exclude non-colonic areas, enabling more precise tumor delineation within the intestine. Model performance was evaluated using the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU). The best performing model achieved DSC and IoU values of 91.1% and 85.7% for colon segmentation, and 78.0% and 69.8% for tumor segmentation, respectively. Results indicate that nnU-Net and YOLO models can effectively perform automated segmentation of the colon and tumors on CT images, with excellent accuracy in colon segmentation. Although tumor morphology and boundaries are diverse, the proposed approach demonstrates robust recognition capability. Furthermore, applying colon masks to assist tumor segmentation significantly improves localization precision and reduces misclassification. This strategy holds promise for integration into clinical decision-support systems to enhance early colorectal cancer detection and treatment outcomes.

ABSTRACT. Early and accurate diagnosis of life-threatening pulmonary diseases like tuberculosis and lung cancer is essential for reducing mortality rates, yet remains challenging. This paper addresses these challenges by proposing a novel deep learning-based segmentation model, EC-CBAM-UNET, that significantly improves the foundational preprocessing step for precise lung field segmentation. The proposed architecture extends the classical U-Net by integrating a modified Convolutional Block Attention Module (CBAM), in which the channel attention component is replaced with an Efficient Channel Attention (ECA) mechanism while retaining the original spatial attention. This directs the network to focus on diagnostically significant regions while suppressing irrelevant features. Furthermore, Efficient Convolutional (EC) blocks replace standard convolutions in both paths, achieving computational efficiency. For hyperparameter tuning of this architecture, a multi-objective optimization technique based on genetic algorithms is used. The resulting model is trained and evaluated on a publicly available medical imaging dataset. Quantitative evaluation demonstrates that EC-CBAM-UNET achieves a Dice Similarity Coefficient (DSC) of 97.4%, a Mean Intersection-over-Union (mIoU) of 95.1%, a Precision of 98.2%, and a Recall of 96.7%, outperforming existing approaches. By providing high-fidelity lung masks, the proposed architecture delivers reliable inputs to downstream Computer-Aided Diagnosis (CAD) models.

ABSTRACT. As the demand for Artificial Intelligence (AI) based solutions continues to grow, AI techniques incorporated into RISC-V(Reduced Instruction Set Computer) architecture have emerged as a compelling solution. The flexibility of the RISC-V instruction set, combined with scalability of AI techniques, proves to be beneficial in hardware accelerator systems for enabling real-time AI applications. In this paper we propose a novel modified RISC-V architecture designed on an FPGA with the addition of NN features providing simplicity in execution and efficiency in performance. Proposed RISC-V processor with its dynamic internal neural network exchange functionality is ideal for external neural computation accelerators in specific applications, such as embedded and edge computing systems. Specifically, this implementation initializes and executes the neural network as a single-instruction operation by utilizing user inputs which are stored in on-chip memory and made accessible to the neural network via RISC-V instructions. The proposed approach is evaluated using MNIST digit classification as a Proof Of Concept. Experimental results show that our approach achieves 7.45 times faster classification compared to the baseline processor computing NN with multiple assembly level instructions. A detailed analysis of processor performance based on Latency, Throughput and Area is presented along with evaluation of accuracy of the classification for different input datasets.

ABSTRACT. Bisimulation is a fundamental concept for analyzing behavioral equivalence in automata and transition systems, with applications in model reduction, verification, and logical reasoning. In fuzzy automata, computing bisimulation relations using classical algorithms such as fixpoint iterations or partition refinement becomes computationally expensive as the state space grows, limiting their scalability to large systems. To address this limitation, we propose Approximate Bisimulation Learning (ABL), the first learning-based framework designed to estimate bisimilarity between states of fuzzy automata. ABL employs a neural architecture combining shared state encoders with a similarity prediction head, trained in a supervised manner on synthetic datasets where exact bisimulation values are computed using classical methods. Once trained, ABL provides fast and scalable inference of bisimilarity, effectively replacing costly iterative algorithms. We evaluate ABL on a dataset of over 10,000 synthetic fuzzy automata with diverse state sizes, transition densities, and fuzzy membership distributions. Experimental results demonstrate that ABL achieves a mean absolute deviation of less than 0.03 from true bisimulation scores while delivering up to 500$\times$ faster inference compared with classical algorithms. These results highlight the potential of learning-based approximate bisimulation for scalable analysis, model reduction, and neural-symbolic reasoning in fuzzy automata and related systems.

ABSTRACT. While generative AI revolutionized content creation, it remained fundamentally reactive—responding to prompts but unable to pursue goals autonomously. Agentic AI emerged to bridge this gap—not through traditional workflow automation's predetermined processes, but through autonomous reasoning that enables AI to pursue goals independently. Unlike workflow automation which executes fixed sequences, Agentic AI can break down ambiguous objectives, make dynamic decisions, use tools adaptively, and collaborate across systems—representing a shift from scripted automation to intelligent agency. However, despite widespread prediction of Agentic AI and the proliferation of frameworks and protocols, enterprise adoption has lagged expectations. This gap stems from a fundamental tension: while agentic capabilities promise autonomous execution, the non-deterministic nature of LLMs creates reliability challenges that compound in multi-step workflows. This talk examines why implementations fail, maps the current landscape of challenges from non-deterministic behavior to cost control, and invites researchers to engage with these open problems—advancing our collective understanding of what production-ready AI agents truly require.