Select Field
Select Ministry
23 results
Electrocardiography (ECG) is a fundamental method not only commonly used in the hospital for clinical requirements but also widely adopted in home and personal healthcare systems to obtain the electrical activity of the heart. An arrhythmia monitoring system (product name: YuGuard) is proposed and used in a clinical trial. The proposed system has three parts. The first is a high-resolution, low-power analog front-end circuit for implementing bio-signal sensing circuits. The features of the circuits are low complexity, high resolution, and low power consumption. The second part is a digital signal processor with a decimation filter and a universal asynchronous receiver/transmitter package generator. The last part is used to realize a software interface on the smartphone for ECG signal recording, display, and classification. A wavelet-based classification method is also proposed to classify the rhythm. The chip used in the system is fabricated through the TSMC 0.18 μm standard CMOS process. The classification algorithm is verified with data from the MIT/BIH Arrhythmia Database. The accuracy of beat detection and arrhythmia classification is 99.4% and 95.83%, respectively. Patients in Tainan Hospital are enrolled in a human study to verify the performance of the proposed arrhythmia monitoring system. Results show that the system can acquire and classify ECG signals.
Future Tech | Biotechnology & Medical careAny defect in a solar cell will reduce the amount of electricity generated. Over time, these defects will result in hot spots. Currently, EL images are commonly utilized for defect detection during the production process of solar modules. However, manual visual inspection is primarily employed formodule defect assessment, leading to inconsistencies in standards and judgments.Our detection technology is based on the original and adaptable Yolo architecture. This technology incorporates innovative approaches like incremental learning and transfer learning to address the challenges associated with diverse flaw types and difficult data collection. Consequently, the accur acy of defect identification has been significantly enhanced, and it can now achieve real-time detection of dozens of defects with an impressive accuracy rate of 99.8%. Furthermore, our solar panel IR thermal defects technology employs a novel object detection method specifically designed for detecting thermal in IR videos. Our team's EL inspection software has been certified by KIWA in the Netherlands. In July 2020, "Msapiens Technology Co., Ltd." was established, and it secured investment from domestic venture capital firms. Presently, venture capital firms from the Netherlands and Germany are showing strong interest. Four Technical Advantages: (1) The largest number: it has more than 21 million EL images and has established the world's largest effective solar module EL image database. (2) The most accurate detection: impressive accuracy rate of 99.8%. Other industries typically achieve accuracy rates ranging from 75% to 95%. (3) The fastest detection rate: Our EL full-load detection speed reaches 18,000 pieces/hour. In comparison, other industries only manage 100 pieces/hour. (4) Highest market share: successfully obtained 8 customers of the world 's top ten solar module manufacture
Future Tech | Green Energy & Environment/Electronics & OptoelectronicsRecent advances in protecting node privacy on graph data and attacking graph neural networks (GNNs) gain much attention. An adversary can utili ze the powerful GNNs to infer users’ private labels in a social network. W e propose to adversarially defend against GNN-based privacy attacks, and present a graph perturbation algorithm, NetFense, to achieve the goal. Ne tFense can simultaneously keep perturbed graph data unnoticeable (i.e., n ot affect social connections), maintain the prediction confidence of target ed label classification (i.e., preserving data utility), and reduce the predicti on confidence of private label classification (i.e., protecting the privacy). E xperiments on 3 benchmark datasets exhibit that the perturbed graphs by NetFense can effectively maintain the performance on targeted label class ification, and significantly decrease the prediction confidence of private la bel classification. The tech is published in IEEE TKDE 2023, and collaborate d with banks E.Sun and SinoPac.
Future Tech | Information & Communications/Machinery & SystemThe average nocturnal blood pressure (BP) or BP dip is a highly sensitive i ndicator of cardiovascular disease risks. However, the nocturnal BP during sleep can not be easily obtained by current BP measurement devices with out disturbing the user’s daily life. As opposed to the point BP estimatio n by smart watch in the sitting position, our research team, composed of r esearchers from National Taiwan University Hospital, National Central University, and Mediatek, uses SENSIOTM smart watch to collect 24-hour pho toplethysmography (PPG) signals together with ambulatory BP monitorin g (ABPM) outputs of the patients for developing a diurnal and nocturnal B P estimation algorithm from wrist PPG in free-living scenario. The applicability of BP estimation during daily life condition is assessed th rough a signal quality indicator established by 3-axis accelerometer and P PG skewness. The correlation coefficient between our activity count (AC) a nd the activity evaluation from sports watch is 0.92 and thus correct judge ment of activity condition can be acquired. To construct a personal model based on multi-layer perceptron and convolutional neural network with fe w training data, the model agnostic meta learning (MAML) technique is a dopted. The stationary data from 535 subjects were employed as the train ing task while the 24-hour data from 14 subjects with age 55.2 23.2 years old were used as the testing task. The query set of testing task contains 6 r ecordings, 3 at daytime and 3 at nighttime. The support set of testing task consists of 5 recordings. The average root mean square error of the BP esti mation per subject is 7.65 mmHg. The mean and standard deviation of tot al SBP estimation error are 0.68 7.98mmHg. The long-term and pervasive monitoring of personal health thus can be achieved by the wearable devic e.
Future Tech | Information & Communications/Machinery & SystemTrianswer is a bio-medical sensor evaluation module designed for wearable and IOT-based products. It not only contains several features about low-power consumption, miniaturization, and IOT-based design but also provides the acquisition of high-quality bio-signals including Electrocardiography (ECG), Electroencephalography (EEG), Electromyography (EMG), Photoplethysmography (PPG), etc. By reducing the development time and cost, developers can manufacture their wearable devices rapidly with the assistance of Trianswer. In addition, acquiring the specific bio-signal needs its corresponding module. Developers can combine the modules like building blocks according to their needs. The main concept of Trianswer is to promote the development of bio-medical wearable devices with different biosensors
Future Tech | Information & Communications/Biotechnology & Medical careAutonomous driving is one of the most important research topics. Deep R einforcement Learning (DRL), a machine learning paradigm that has made remarkable success, say mastering Go and chess without human knowled ge like AlphaGo and MuZero, has emerged as one attractive method for a utonomous driving. By merging with computer vision, DRL can control ve hicles directly through the front camera, known as end-to-end control. Thi s makes the approach to be highly adaptive to diverse environments and r oad conditions for research. First, we propose an image-based sim-to-real transfer technique that redu ces performance drop, when transferring DRL models from virtual to real environments. By using CycleGAN, we transform simulator images to a re al-world style and train the model accordingly. This approach substantiall y minimizes the sim-to-real gap in pure vision models, greatly enhancing adaptability for self-driving race cars on real tracks. Furthermore, our research highlights the importance of control smoothne ss for driving stability in self-driving cars, and thus proposes a new metho d for action smoothness. In contrast, the traditional DRL methods focused on maximizing cumulative rewards, which often make actions jerky and u nstable. In our new method, we propose to use a loss function to ensure c ontinuity in consecutive actions, improving the smoothness of actions an d significantly enhancing control stability. Moreover, we also observe that smoother actions also lead to increased race car speed. After overcoming these challenges, our team has developed effective and stable DRL policies that greatly enhance the completion rate and reduce c ompletion time for physical cars. This technology was successfully showca sed at prestigious workshop events during ICRA 2022 and IJCAI 2022 conf erences. Furthermore, in late 2022, our achievements stood out by winnin g the top three in the AWS DeepRacer League competition hosted by Ama zon, with 150,000 contestants.
Future Tech | Information & Communications/Machinery & SystemProspects: MEMS micro scanning mirrors are one of the core technologies for realizing LiDAR, HUD, HMD, and smart glasses. Due to their competitiv e advantages in terms of size, power consumption, and cost, MEMS micro scanning mirrors are considered a highly promising key technology. For th e abovementioned applications, MEMS micro scanning mirrors must meet the challenges of design specifications such as large angles, high frequenc ies, and large mirror sizes. To meet the upcoming market and business op portunity, different countries form so-called multi-national R&D alliance one after another. For example, “LaSAR”, led by STmicroelectronics, con sists of several European technology power (Germany, France, Italy, and s o on). Although there are also Taiwanese companies involved, the key tech nology of MEMS micro mirror still cannot be hold in hand. Goal: (1) Component: We have developed piezoelectric MEMS micro scan ning mirrors that meet specifications for automotive and consumer electr onics applications. (2) Process: Through collaboration with foundries, we h ave achieved commercialization of the piezoelectric process platform and assisted in establishing a database of piezoelectric parameters. (3) Modul e/Applications: The fabricated MEMS micro scanning mirrors are integrate d into LiDAR and HUD modules based on specific usage requirements, en abling the integration of optical, mechanical, and electrical modules/syste ms. Features: (1) Mirror size (LiDAR): With a mirror diameter of 3 mm, the scanning angl e can reach 140 degrees. (2) Operating frequency (Display): At a driving frequency of 37 kHz, the sc anning angle can still reach 70 degrees. (3) Scanning angle (LiDAR / Display): Achieving a scanning angle of 180 de grees without the need for vacuum packaging. (The world's first) (4) Operating mode (Display): Capable of projecting nearly vertical patter n boundaries in dual-axis imaging.
Future Tech | Information & Communications/Machinery & SystemHidden hearing loss is a challenging condition to detect early on, often lea ding individuals to seek treatment only when their hearing is significantly impaired, resulting in irreversible consequences. Our system comprises fo ur main components:1.The "Ear Scale App": This application serves as a hearing screening tool,which when used with specialized headphones, allows users to conduct he aring screenings in non-soundproof environments.2.An intelligent big data application system: This system utilizes big data t echnology to analyze and compare the collected data, enabling statistical analysis and trend predictions for deeper insights into hearing loss patter ns. 3.A machine learning cloud monitoring system: Leveraging the data from the intelligent big data application system, this component establishes m achine learning models to monitor changes in users' hearing abilities ov er time. By detecting early signs of hidden hearing loss, the system can pr ompt timely intervention and treatment to prevent further deterioration. 4.Active noise-cancelling algorithms: Integrated into the headphones, the se algorithms enhance the accuracy of the hearing tests by eliminating ext ernal interferences, ensuring precise results even in various noise scenario s. We have validated the feasibility of this system through a pure-tone scree ning activity conducted at Tianmu Elementary School in Taipei City. The ex perimental results indicate that our system not only detects hidden hearin g loss in children at an early stage but also produces comparable results t o those obtained from professional hearing testing instruments. In summary, we envision that through this system, individuals with hidden hearing loss and sudden deafness can perform self-assessments at home, track their hearing status, and receive comprehensive assistance for audit ory rehabilitation.
Future Tech | Information & Communications/Machinery & SystemYuBelt is a belt with a real-time ECG signal monitoring device, which can be patched on the chest for sportsmen. It can be a good partner of Your「heart」to detect the rhythm of the heart respond to the health status and provide the calorie and physical fitness assessment.
Future Tech | Information & Communications/Biotechnology & Medical careThrough combining pet ECG acquisition and a front-end sensor, the I-Pet ECG patch is a wireless bio-signal detecting system for acquiring electrocardiogram signals from a pet. An algorithm is implemented to evaluate the emotional state and heart rate variability. Based on these analysis results, this system can be a reference for veterinarians and provide suggestions to users. Moreover, the emotion index can help users to understand the mental states of their pets. This work combines the concept of wearable devices and the Internet of Things and contains the hardware implementation and software development. With a software application such as a graphical user interface, the user can obtain the real-time and past condition of the pet. This information platform is a channel to communicate with the vet and a bridge for pet lovers to exchange their experiences.
Future Tech | Biotechnology & Medical care/Information & CommunicationsAs we all know, road surfaces in Taiwan are often structural failures, not on ly cause discomfort, but also affect ride comfort, and pose a direct safety r isk, where an impact could initiate suspension structural damage, leading to a tire blow out later. In addition, Gartner predicts that the total number of autonomous-ready vehicles worldwide will exceed 740,000 by 2023, of which nearly 60% will be mainly for public transportation connections. He nce, in addition to vehicle safety, the ride comfort is an important issue of self-driving car. Our teams consists of six professors and working with five automotive industries to jointly develop a domestic 『Intelligent Electroni c Suspension (IES) System』. By combining AI technologies, the IES syste m can promote ride comfort and vehicle stability, and increase its durabilit y. - Taiwan Pavement Defect Image Dataset (TPDID) The world's first dataset for TPDID has been established, which the total number of TPDID samples is 89,723. The TPDID dataset have more sample s, multiple types of road defects, and conforming to the actual road condit ions in Taiwan, which can not only improve the accuracy of the developed AI-based road defects perception system, but also effectively increase roa d repair efficiency by sending the road defect detection results to the road way maintenance unit in real time. - Intelligent Electronic Suspension (IES) system Nowadays, most of the vehicle suspension systems are composed of tradi tional rigid spring and fixed damper, only a few high-priced top vehicles h ave an active suspension system; however, it is still limited to manual adju stment. By combining AI technologies, the IES system not only promote ri de comfort and vehicle stability when passing the road defects, and increa se its durability. In addition, the lightweight AI and embedded technologi es have entered the automotive supply chain, Nissan commercial vehicles, together with Chimei, and the production value will expect NTD 750 mi
Future Tech | Information & Communications/Machinery & SystemPublic touch panels like those on elevators and ATMs could serve as medi ums for virus transmission, an often-overlooked risk in epidemic preventi on. Current preventative measures have significant drawbacks: Disinfection of protective films is labor-intensive and wasteful. Voice recognition alters user behavior and has concerns over droplet trans mission. Facial recognition requires upfront labor and isn’t widely applicable. Infrared sensors are unresponsive and necessitate wasteful replacements. Our team developed "m’AI Touch," an Motion-Based AI Contactless Tech nology, combining various sensors and algorithms to accurately determin e a passenger’s intended button selection. It enables users to maintain t heir habits while reducing the risk of virus transmission. The technology is primarily intended for places like hospitals and quarantine hotels, with sig nificant foot traffic and virus transmission risk. With the support of Tsinghua University, we’ve conducted tests on camp us elevators and received orders from national mask companies. We’re p reparing for tests with hospitals and companies in Hsinchu Science and In dustrial Park , hoping to reduce virus transmissions such as enterovirus, n orovirus, and Covid-19 significantly. Our technology, akin to external access control devices, is suitable for bot h new and old elevators of any brand. We aim to cooperate with builders, communities, and building owners, targeting commercial buildings, gover nment agencies, large communities, and small residential buildings. Our g oal is to install and set up contactless AI epidemic prevention elevators, be nefiting health and welfare in society.
Future Tech | Information & Communications/Machinery & System1. Technical Background: Critically ill patients require timely medical intervention to prevent irreversible consequences. Supported by the "Artificial Intelligence Special Research Program for Addressing National Major Challenges" by the National Science and Technology Commission, we've developed advanced time-series machine learning technologies, including multi-objective early prediction techniques and an intensive care alarm system. These innovations utilize deep reinforcement learning and multi-objective optimization with various physiological signals to construct models and systems. Our system addresses the challenge of achieving accuracy and earliness in intensive care, transforming passive healthcare into an active intervention model. 2. Technical Content: The early warning system comprises the following components: 1) Segmental Policy Networks, used for processing time-series physiological signals and extracting features from various segments; 2) a Knee-Guided Neuroevolution Algorithm for multi-objective optimization to balance accuracy and earliness; 3) Constraint-Based Knee-Guided Neuroevolution Algorithm, allowing prioritization of prediction targets based on the scenario; 4) a Control Agent Module that employs reinforcement learning techniques to control the decision-making process; and 5) a Discriminator for outputting prediction results. Ultimately, a comprehensive intensive care alarm system is composed of the aforementioned modules. 3. Technical Overview: Our system has undergone testing on various large-scale medical time-series datasets. The results indicate that it significantly outperforms state-of-the-art methods. We have published four papers in top international journals: IEEE TKDE(IF=9.235), IEEE TNNLS(IF=14.255), and IEEE JBHI(IF=7.021). The technology has successfully transitioned to hospitals and undergone clinical field trials. Additionally, it has received commendation on the AICoE website.
Future Tech | Information & Communications/Machinery & SystemAn electroless plating method is, for the first time, adopted to synthesize a nickel-sulfur (Ni/S) energy-storage material for the development and co mmercialization of the third-generation rechargeable batteries, which is t he high-energy-density lithium-sulfur cells. The patented nickel-sulfur en ergy-storage material has a layer of conductive nickel nanoparticles wrap ping the insulating sulfur particles to accelerate the electron transfer and t he reversible electrochemical reaction, while decelerating the loss of activ e material during the long-term conversion reaction. As a result, our nicke l-sulfur energy-storage material enables the lithium-sulfur cells to simulta neously attain a superior high amount of sulfur approaching 60-95 wt% a nd 10-14 mg cm-2 with a lean-electrolyte conduction featuring an electrol yte-to-sulfur ratio of just 7 μL mg-1, yet exhibits high energy density of 13 –28 mW∙h cm-2, long cycle life of 200 cycles, and excellent rate performan ce of C/20-C/2.
Future Tech | Information & Communications/Machinery & SystemOur project presents an integrated and scalable precision health service fo r health promotion and chronic disease prevention. Continuous real-time monitoring of lifestyle and environmental factors is implemented by integ rating wearable devices, open environmental data, indoor air quality sensi ng devices, a location-based smartphone app, and an AI-assisted telecare platform. The AI-assisted telecare platform provided comprehensive insig hts into patients'clinical, lifestyle and environmental data, and generate d reliable predictions of future acute exacerbation events. This service has been used in many hospitals (NTUH, ECKH, FJCUH, TSGH, CTH, MMH, and Okayama University Hospital) in Taiwan and Japan. Through content mod ular design, it can provide health management services for patients with d ifferent chronic diseases. All data from ~2000 patients were collected pros pectively during a 24-month follow-up period, resulting in the detection o f 416 abnormal episodes. ML and DNN were used to train modular chroni c disease models. The modular chronic disease prediction models that hav e passed external validation include obesity, panic disorder, colorectal can cer and chronic obstructive pulmonary disease, with an average accuracy of 85%, a sensitivity of 76%, a specificity of 86%, and an F1 score of 81%, r espectively. Compared with previous studies, we establish an effective way to collect li festyle, life trajectory and symptom records, as well as environmental fact ors, and improve the performance of the prediction model by adding obje ctive comprehensive data and feature selection. Our results also demonstr ate that lifestyle and environmental factors are highly correlated with pati ent health and have the potential to predict future abnormal events better than using only questionnaire data. Furthermore, we have constructed a c ost-effective model that needs only a few features to support the predicti on task, which is helpful for deploying real-world modular
Future Tech | Information & Communications/Machinery & System1. Low-power intelligent image sensing system: Develop low-power image sensing chip design with time information calc ulation to provide low-power image sensing for artificial intelligence appli cations. It can be applied to dynamic object detection to achieve temporal information acquisition and fast processing, effectively reducing the hard ware complexity of the back-end computing unit. Using low-power time i nformation to calculate the front-end computing capability in the image s ensor chip, to achieve the front-end analog frame difference dynamic resu lts, not only reduce the amount of data transmission between chips but al so provide a high-efficiency, low-latency visual sense of neural network ar chitecture test solution. 2. Obstacle detection and obstacle avoidance neural network: Through the analysis and exploration of fruit fly vision, the bionic optical f low compatible with the front-end calculation of the image sensor chip, th e low-resolution dense depth sensing of monocular vision, and the simple and effective obstacle avoidance logic have been developed, and the fram ework for the entire system has been formulated, and aiming at visual sen sing, chip integration, etc., to design and develop a low-power, real-time, i ntelligent vision system with recognition functions. 3. Configurable population-based spiking neural network processor: When designing, developing, and integrating neuro-imitative system chip s on miniaturized devices, it is hoped to achieve real-time obstacle avoida nce with low power consumption. The event-driven feature is suitable for matching with the dynamic sensory data of machine vision. The neural gr oup spike neural network we developed is responsible for behavior decisi on-making to achieve the effect of obstacle avoidance. The neural group f ormed by eight neurons represents eight three-dimensional Different dire ctions in space (front, back, left, right, up, down, clockwise, counterclockw ise).
Future Tech | Information & Communications/Machinery & SystemIn recent years, the product life cycle has become shorter, and sometimesit is necessary to obtain a pre-order within a month by trial production and shipment, then mass production. However, in the engineering sample validation stage, we are faced with at least 30 pcs of mechanical parts and appearance parts for initial design validation testing and then 100-1000 pcs of small-lot trial production validation. The problem is that it is easier to finish quickly or ship with design verification if you want to make more than 30 pcs of engineering samples, whether by CNC machining or directly with sample molds. Even if we achieve the target, the product risk is exceptionally high, the cost is incredibly high, and the efficiency is extremely low. Currently, the following four techniques use for engineering prototyping, including RP rapid prototyping, 3D printing, CNC machining, and plastic injection molding. The portable micro injection machine with IoT intelligent monitoring developed by this technology can solve the product and technology gaps in the industries, as mentioned earlier, including easy-to-obtain materials, parts with complex appearance, structural strength, long-term storage, rapid small-lot production, cost reduction, no secondary processing of details before and after processing, and inexperienced operators. The following four futures as small batch production, pre-volume production evaluation, po rtable design, and all-around application.
Future Tech | Information & Communications/Machinery & System/Life ApplicationPPFHIR (Privacy Preserving Fast Healthcare Interoperability Resources) isa brand new system for exchanging electronic healthcare information. It is b ased on the FHIR standard and various advanced information security tec hnologies to protect personal medical data on the public cloud and help medical institutions reduce the cost of maintaining IT systems. We used S ABHPRE (Searchable Attribute-Based Homomorphic Encryption with Prox y Re-Encryption) to encrypt the medical data to avoid inner attackers (pub lic cloud) prying about the data. It can also keep the encrypted data havin g more flexibility than traditional technologies. With SABHPRE, the medic al institutions can ask the public cloud to search, aggregate, and re-encry pt the encrypted data. Our system also contains Federated Learning which can help the medical institutions cooperate with foreign medical institutio ns to build machine learning models but never leaking any patient’s priv acy information.
Future Tech | Information & Communications/Machinery & SystemComing soon!