Research Topic

I am interested in VLSI testing, defect diagnosis, AI design and IoT system integration. In my first 10-year career from 2007 to 2017, my research topic is focused on semiconductor including IC design, VLSI testing and defect diagnosis. After 2018, I pay more attention on AIoT co-design and integration, and make all-out effort to lead HysonTech team to develop and launch several AIoT commercial products, and has been successfully used in aquaculture, smart traffic enforcement, manufacturing automated inspection, medical diagnosis, biotechnology analysis, marketing prediction and other fields. Starting in 2023, I established and led a circular economy technology development team and began to invest in the development of circular economy technology and related materials to achieve various United Nations Sustainable Development Goal (SDG). 

The following brief introduction is about the following topics I involved:

1. Semiconductor Field (2007-2017)

• Logic Build-In Self-Test

• Test Response Compaction

• Low Power Testing

• Wafer-level Defect Diagnosis

2. AI and IoT Field (2018-present)

• AIoT design on Intelligent Aquaculture

• AI Traffic Enforcement

• AI-assisted Rapid Test for Gerontology

• AI Technology for Marine Conservation and Safety of Water Activities

• AI Price Prediction

• AI Image Enhancement: Demoireing, Defog, Color Restoration

• AI for Tiny Object Detection and Classification: Mites for example

• AI Quality Classification for Metal Thin Film

• AI Identification and Comparison for Counterfeit Corporate and Government Seals

3. Circular Economy Field for Sustainable Development Goal (2023-present)

• Low Carbon Smart Electronic Material System for P/N Monitor and Removal

Background of VLSI Testing

In today's industry, Automatic Test Equipment (ATE) is being used to quickly confirm whether a manufactured design works or not. Figure 1 shows an example to use ATE for evaluating a design-under-test (DUT). The ATE is responsible to deliver test patterns to the inputs of DUT, and compare the output test responses with the fault-free responses, i.e. the golden responses which are stored in ATE. All discrepancy between the output responses and golden responses will be recorded into the final log files for further analysis. In this way, these faulty devices can be filtered out before entering the market. To accomplish this method, the test data including all required test patterns together with the associated fault-free responses must be stored in ATE for complete testing.  

However, both the I/O bandwidth and the memory size in ATE are limited. When we need to test a very complex design, such as system-on-a-chip (SOC) design that integrates many cores into a single chip, the required test data volume will rapidly increase, while a very longer test application time is also needed. This will induce an extremely high test cost of ATE, and also delay the time-to-market. The report of 2013 International Technology Roadmap of Semiconductors (ITRS) predicts that the worst case of the test data volume for a customer SOC design will grow to 19886Gbits by the year of 2025. In 2025, the test data stored in ATE needs to be compressed to 1/1376 of the original data volume in order to maintain scaling of test costs and product quality. Thus, how to efficiently reduce both the test pattern and response data volume has become a critical problem in both academia and industry. It is also listed as a serious bottleneck in future SOC design.

Logic Build-In Self-Test

Logic built-in self-test (BIST) techniques that embed some specified test infrastructure into a DUT to test the device itself, is widely used as an effective way to reduce test cost. In this topic, we develop an internal-response-feedback self-test technique to achieve complete testing in a very short time with no any storage requirement. Since all required test patterns can be on-chip generated, the storage test cost for input test data can be zero. To our knowledge, this technique is the fastest technique so far among all methods in the world that can complete test all ISCAS benchmark circuits without using any internal or external storage device.

Extensive Reading:

1. W.-C. Lien and K.-J. Lee, “A Complete Logic BIST Technology with No Storage Requirement,” IEEE Asian Test Symposium (ATS), 2010, pp.129-134.

2. W.-C. Lien, K.-J. Lee, T.-Y. Hsieh and W.-L. Ang, “Efficient On-Chip Test Generation Scheme Based on Multiple Twisted-Ring Counters,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (T-CAD), Vol. 32, Issue 8, pp. 1254-1264, Aug. 2013. 

Test Response Compaction

Unlike the input pattern compaction, test response compaction techniques have to solve the problems of error aliasing, unknown-value, and poor diagnosability during compaction. To solve these problems, we investigate an output bit selection technique for test response compaction. This technique has several advantages including high compaction ratio, zero aliasing, fully X-tolerance, high diagnosability, low area overhead and simple test control. Also no circuit/ATPG modification is needed, hence this method can be easily integrated into any typical industrial test flow to significantly reduce the test cost of both DC- and AC-scan testing with no pattern inflation and no test quality loss.

Extensive Reading:

1. K.-J. Lee, W.-C. Lien and T.-Y. Hsieh, “Test Response Compaction via Output Bit Selection," IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (T-CAD), Vol. 30, Issue 10, pp.1535-1544, Oct. 2011.

2. W.-C. Lien, K.-J. Lee, T.-Y. Hsieh, K. Chakrabarty and Y.-H. Wu, “Counter-Based Output Selection for Test Response Compaction,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (T-CAD), Vol. 32, Issue 1, pp. 152-164, Jan. 2013.

3. W.-C. Lien, K.-J. Lee, “Output Bit Selection Methodology for Test Response Compaction,” IEEE International Test Conference (ITC), 2016, pp. 1-10. 

Low Power Testing

In addition to reduce the overall test cost, test power is also needed to be considered. During test application, test power may be significantly higher than the normal functional power because circuits may enter nonfunctional states for testing. Excess test power may slow down a normal circuit and eventually fail the test, thereby resulting in test-induced yield loss. In this topic, we proposed a power-safe test pattern determination procedure to determine a set of power-aware test patterns that will not violate any power constraints and still maintain the same test quality. It is worth to mention that this method has been proven to completely solve excessive test power issue without or with less additional test vectors. Actually, in many circuits especially for the larger circuits, the number of test vectors is even smaller, which means that the test cost can be maintained at the same or even less. Therefore, our technology has been widely cited and used in the test process by industry and academia.

Extensive Reading:

1. Y.-H. Li, W.-C. Lien, I.-C. Lin and K.-J. Lee, “Capture-Power-Safe Test Pattern Determination for At-Speed Scan-Based Testing,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (T-CAD), Vol. 33, Issue 1, pp. 127-138, Jan. 2014.

Defect Diagnosis

After testing, fault diagnosis process is used to predict the defect locations and types in failing chips. The information obtained through this process can be used to guide the destructive physical failure analysis process to point out the actual defect locations and ascertain the causes of the defects. This is a critical procedure for IC manufacturing companies to ramp up the yield of new process technology. One important step of fault diagnosis is to generate a suitable set of patterns that can be applied to the chips via ATE such that proper diagnosis information of the chips can be collected and analyzed. In this topic, we developed efficient diagnostic pattern generation procedures to generate required diagnostic patterns to reach almost 100% successful physical failure analysis on manufactured designs. To the best of our knowledge, this is the first reported work that can achieve full diagnosis resolutions for ISCAS and IWLS benchmark circuits, and also on real manufacturing defect. This technique has been patented and integrated and used in all TSMC’s advanced processes.

Extensive Reading:

1. C.-H. Wu, K.-J. Lee and W.-C. Lien, “An efficient Diagnosis Method to Deal With Multiple Fault Pairs in a Single Circuit,” IEEE VLSI Test Symposium (VTS), 2014, pp. 240-245.

 HysonTech FarmABetterFish Intelligent Aquaculture System

Smart farming fisheries is a global trend. As the population continues to grow, the demand for protein will also increase. Eating fish is the healthiest, but the fishery resources that the ocean can provide are limited. In contrast, aqua-farming is more sustainable for the environment than the wild-capture. However, the aquaculture industry in the Asia-Pacific region has long suffered from the negative influence and industrial predicament of high risk, high pollution, and high dependency on manual labor. For example, in Taiwan, the aquaculture industry has been unable to meet the demand of the domestic market, and can only rely on the imported catches.

To make aquaculture easier, Hyson Technology team utilized AI and IoT technologies, to propose an innovative intelligent aquaculture management system, called the "FarmABetterFish System", to help the world achieve the goal of sustainable aquaculture production. We have successfully developed an AIoT-based technology to monitor, measure and control the growth status of fish and shrimp. FarmABetterFish system is estimated to automate 75% of routine works, reduce 40% of risk, and elevate 30% of production growth of the current aquaculture industry. Originated from a user-centric perspective, the system incorporates both technological and business model innovation, making it the best practice of inclusive business. It is the best product known so far in the world to move towards sustainable production of aquaculture. There have been many landing applications that the AI-powered system helps the local fishermen on improving the manpower efficiency, judging the pool division, planning the shipment, etc. I believe it is an exemplary example of traditional fishing industry undertaking a process of digital transformation.

The AIoT-Based FarmABetterFish system is a joint collaboration of various stakeholders, including local aquaculture industry, the government sector, universities, and research institutes. The interconnected alliance helps create and optimize the full potential of the FarmABetterFish system. This system is used to reach the sustainability of local fish farming industry, significantly reduce the risks and burdens at work, attract young people to participate in the aquaculture industry, and substitute fish farming for capture fisheries, thus protecting the marine environment and resources while reducing fish feeding waste, the water and marine pollution. This product and aforementioned actions have high correlations with the following four Sustainable Development Goals (SDGs) proposed by the United Nations: Zero Hunger, Decent Work and Economic Growth (SDG#8), Responsible Consumption and Production (SDG#12), and Life Below Water (SDG#14).

FarmABetterFish system can benefit over 40,000 families associated with the aquaculture industry in Taiwan, improve the daily life of over 20 million fishermen around the world, and produce sustainable and environmental-friendly protein for human society. 

Visit HysonTech Website: https://www.hysontech.tw 

 HysonTech AI Traffic Enforcement System

Nowadays, the police unit is responsible for complex and heavy daily tasks and duties, including patrols, DUI checkpoints, regular inspections, and standby duties. In terms of traffic law enforcement, in addition to dealing with traffic accidents or violations on the spot, it is also necessary to deal with traffic violation video reports. Since the number of such traffic violation videos is increasing year by year, it takes a lot of workforce and time to check the footage, find the violation clips, locate the license plates, judge the violation type, and calculate the driving speed before the collision. In order to reduce the loading of police unit, we have developed two systems, one for Intelligent Traffic Violation Detection and one for Car Accident Analysis and Driving Speed Estimation, as described below:

1. Intelligent Traffic Violation Detection System: HysonTech intelligent traffic violation detection system which can analyze traffic violation video to detect and classify the traffic violations, and also calculate the driving speed before the collision if an accident happened. Besides, this system can automatically generate fines and a court indictment in seconds. This fully-automated product can greatly reduce the need for human judgment or possible human misjudgment so as to reduce the police's complicated tasks and bring more peaceful and secure society. It is currently used in the regular traffic law enforcement system of the Kaohsiung City Police.

2. Traffic Accident Analysis and Driving Speed Estimation System: In the past, when the police handled traffic accident cases, they often needed to consult intersection monitors and driving recorder images provided by the public, and repeatedly viewed and analyzed the images to calculate the driving speed in order to clarify the cause of the accident. On average, each police officer spends 40 working days per year performing this duty. It takes a lot of manpower and time. To solve this problem, HysonTech team cooperated with the Traffic Police Corps of Kaohsiung City Police Department to develop the world's first "AI-Assisted Traffic Accident Speed Estimation System". This system supports two methods commonly used by the police: Grid Method and Cross-Ratio Theorem. The AI-assisted speed calculation of the vehicle at the time of the accident is based on the input images from the intersection monitor and driving recorder to automatically generate reports that can be used in court. Tests in hundreds of real cases have shown that this system is not only highly accurate but also fast in calculation, and can reduce police time by at least 80%. This system is the First Scientific Forensic Evidence in Taiwan to use AI-assisted computing, and it was awarded the 2024 National Police Award, the highest honor awarded by the Police Department.

Contact Us for more product information: https://www.hysontech.tw

 AI-assisted Rapid Test to identify Early Signs of Parkinson’s Disease 

HysonTech team has cooperated with several large medical institutions to present a cloud-AI-assisted system that aims to facilitate the identification of potential Parkinson's symptoms in a fast and simple manner. The proposed approach involves taking a picture of hand-sketched spiral graphics using cell phones by the individuals, which is then uploaded to the cloud for AI diagnosis. Various AI detection models have been applied to consider spatial and frequency characteristics of the spiral graphics. According to the blind test data of hundreds of people, both the sensitivity of Parkinson's disease patients and the specificity of healthy people exceeded 90%. It's also worth mentioning that the success rate of early rapid test for the 1st-phase and 2nd-phase Parkinson's symptoms is 91% under blind test.

This system demonstrates a potential solution to the medical predicament of early-stage Parkinson's disease identification, which often goes undiagnosed or late-diagnosed due to patients' reluctance to accept the condition. Additionally, this study highlights how the impact of the restriction on medical resources in rural areas in the post-COVID-19 era has further exacerbated the problem of accurate Parkinson's diagnosis. Overall, the proposed cloud AI-assisted system presents a promising solution for enhancing the early detection and management of Parkinson's disease anytime and anywhere. HysonTech team is currently working with several hospitals to landing this system, including large medical centers, regional hospitals, and district hospitals, etc. In near future, similar techniques will be used in more geriatrics and elderly disease care, such as Alzheimer's disease.

Contact Us for more product information: https://www.hysontech.tw

 AI Technology for Marine Conservation and Safety of Water Activities

HysonTech and Keelung Marine Science Museum jointly developed the world's first underwater sandstorm AI monitoring and early warning service, which was launched in the Tideland Marine Conservation Area in Taiwan's diving holy land. It combines 5 AI models to analyze 40 types of tidal radar station data and underwater monitoring images. With various data such as ocean sounds, it can detect >90% of underwater sandstorms and provide early warning at any time to ensure the diving safety of divers, including researchers and tourists, and accurately predicts 72% of future sandstorms. Now you can add friends on LINE to try it out at any time anywhere:

(1) Check current and past 3 Hours sea conditions.

(2) Check the probability of sandstorm occurring in the Next 6 Hours.

(3) Automatic early warning and reminder immediately when sandstorm occurs.

This service not only ensures the safety of divers, but also helps to understand coral restoration research, creating a win-win situation for diving researchers, tourists and coral ecological conservation.

In addition, HysonTech team has further developed AI fish species auto-identification, allowing AI to identify and classify >70 precious coral reef fish species through underwater cameras, with an identification accuracy of 90%, and established a cloud-based AI fish species identification and classification platform. Not only the conservation database can be continuously accumulated, but after the identification is completed, the statistics and record reports required by the marine science museum will be auto-generated. This platform significantly reduces >70% of the human resources spent on statistics, records and report writing in the past.

To understand the effect of coral reef restoration, the marine researchers and diving volunteers need to regularly observe and count fish species on the seabed. However, due to the wide variety of coral reef fish, counting is difficult and it is difficult to effectively understand the true recovery situation. Our platform allows researchers to identify and count fish species based on cloud AI. The results of the classification are used to help the marine researchers formulating more effective coral reef restoration strategies.

Contact Us for more product information: https://www.hysontech.tw

 AI Price Prediction

The aquatic market price is an essential indicator for fishermen in making decisions regarding conveyance, fish-catching, and wholesale strategies. As a guide for the aquaculture industry, understanding market prices, the geographical distribution of fish prices, and the prediction of fish prices are vital. In Taiwanese fisheries, however, predicting aquatic prices is challenging due to their drastic fluctuations caused by tropical and subtropical climate variations, export/import fish quantity changes, and political and economic situation uncertainty. To address the non-linear features of aquatic product prices in Taiwan, we proposed a hybrid network model comprising six neural algorithms for aquatic product price forecasting in Taiwan. We have developed predictive models using data from the past 21 years across 14 major wholesale markets by utilizing open transaction data from the Ministry of Agriculture (MOA) and selected variables. The overall average prediction accuracy exceeds 90%, which can even further improve to over 94.27% as the insufficient training data (less than 2000) were excluded.

On the market side, we have developed the SmartFishery aquatic product price prediction service, a fully automated self-learning AI system to acquire data and regulation automatically for a more accurate fish price prediction. Furthermore, we have integrated this SmartFishery service into the LINE app for custom-built operation. Looking ahead, we plan to expand the analytical capacities and consumer-based operations of the SmartFishery system. Additionally, we aim to include a broader range of aquatic products for more comprehensive fish price forecasting. We envision that this developed aquatic prices prediction service will contribute to increased profitability in the aquaculture industries and serve as a scientific reference for the government in formulating relevant aquacultural policies for sustainable fishery management.

Recently, the SOP analysis technology used in fish price prediction has also applied for patents and has been commercially used by customers in many different fields to predict different subject matter such as oil, rubber, steel and stock markets with an accuracy of over 90%. This AI price prediction service has also achieved good market sales results.

Contact Us for more service information: https://www.hysontech.tw

 AI Image Enhancement: Demoireing, Defog, Color Restoration

Taking photos of a digital display or an object with repetitive delicate patterns often causes the images to have unwanted rainbow-like visual artifacts, degrading image quality. Since the pixel arrangement of the display or the object pattern photographed does not agree with the sensor arrangement of the camera, the two arrangements are superimposed in forming the color and shape irregularities, called Moir´e patterns. Unlike other image restoration tasks, the difficulty in eliminating Moir´e patterns is that they appear in a wide range of frequencies with irregular shapes and rainbow-like colors. We propose a Multi-Scale Fusion Network that contains dilated-dense attention, multi-scale feature interaction, and multi-kernel strip pooing. These designs can better catch Moir´e patterns, extract authentic image features, and restore Moir´e images. In addition, we propose augmenting training data by transferring Moir´e patterns to clean images, enhancing demoireing performance as shown as the figures below. The results show that our AI model performs favorably against state-of-the-art demoireing methods on real datasets. This technology has been used in the smart manufacturing processes of the world's largest panel factories and chip industry manufacturers to significantly improve the AOI defect detection performance.

In addition, we also use adversarial network models combined with image signal processing algorithms to perform image defogging and color restoration, and restore images for underwater and land images to improve their visual quality and performance of computer vision as shown below. Related methods have been patented and commercialized for sale. It is also used in our FarmABetterFish intelligent aquaculture product, diving photography, offshore wind power exploration, military underwater exploration, land field monitoring, etc., and is also integrated into the remotely operated underwater vehicle, i.e. ROV robot.

Contact Us for related product information: https://www.hysontech.tw

 AI for Tiny Object Detection and Classification: Mites for Example

From the random sampling of Taiwanese families, we found that in average >2,000 mites are hidden in every gram of indoor dust. Excessive mites often cause allergies in people, especially children. It induces that the number of allergic children in Taiwan has always been high. In the past, it is very difficult to judge the status of mite contamination. Therefore, HysonTech team is cooperating with Japanese biotechnology companies to develop mite-capture technology and use AI technology to quickly and accurately assess the level of mite contamination in each environment and give warnings so as to reduce people's nasal allergies, asthma, atopic dermatitis and other symptoms caused by excessive mites. Our mite species identification and counting service combines 3 AI modules to detect and calculate the number of captured dust mites and carnivorous mites. The accuracy of classification and identification exceeds 85% compared with the time-consuming manual inspection which needs to count through a microscope. The error in the laborious calculation is only 7%, which can more efficiently help people accurately assess and reduce dust mite contamination in their homes. In addition, we have integrated 3 different image enhancement technologies to greatly improve the visibility of tiny mites. Based on our image enhancement techniques, the previously invisible mites can be enhanced to a level that can be recognized by humans. This also greatly improves the accuracy of AI identification. Our techniques allow both humans and AI to see each mite clearly. After this product service is launched in the near future, it is expected to provide to supermarkets, bedding stores, cleaning industries, pest control industries, medical institutions, and the general public to assist the public in preventing and controlling the mite contamination, especially for the families with allergic children. This AI service will be practical and thoughtful, and very close to people’s normal life.

AI Quality Classification for Metal Thin Film

In the manufacturing process, semi-finished products (such as panel films) are often classified according to the results of different test pieces to determine subsequent processing methods and products that can be produced and applied. For example, T0 grade has low resistance and the best coating quality. The semi-finished products can be used in the manufacture of solar panels, while the second-best T1 grade can be used in electromagnetic wave shielding, the T2 grade can be used in thermal temperature control panels, and the T3 grade with the lowest coating quality can be used in touch sensor manufacturing.

Traditional test piece quality classification method requires contacting the test piece with a high-precision probe to measure its resistance value, and classifying the plating quality based on the measured resistance value to determine the subsequent process. However, this contact measurement method will not only cause damage to the test piece, but may also affect the subsequent processing of the product. In order to solve this problem, we have developed an AI technology that uses AI image recognition technology to achieve non-contact plating quality inspection.

We use a large number of metal film images, their corresponding resistance values, UV-vis spectral curves and traditional method classification results as data to train the AI, allowing the AI ​​to learn to classify film quality through image analysis. Currently, AI detection and classification are accurate The degree can reach more than 80%.

This method can not only avoid the damage caused by traditional probe measurement methods, but also use a simple mobile phone lens to perform non-contact AI test piece quality identification and classification, greatly improving measurement efficiency and applying it to automated production lines. This invention will bring new applications and detection technologies to many manufacturing industries.

Contact Us for related product information: https://www.hysontech.tw

AI Identification and Comparison

for Counterfeit Corporate and Government Seals

Seal images are usually easy to gain the trust of the public. Therefore, in recent years, when many fake prosecutorial agencies and investment consulting companies are defrauding, scammers often use fraudulent documents stamped with forged seals to gain the trust of victims. The police colleagues who rushed to the scene to stop the fraud were unable to determine the authenticity of the seal in a short time. They had to send it back to the police station for further comparison, making it very difficult to stop the fraud and collect evidence. In order to solve this problem, the HysonTech team has developed an AI system to quickly identify forged seals, allowing the police to take photos and upload images of suspected forged seals at any time through police handheld computer devices or mobile phones, and the AI ​​will automatically compare them with the police database of forged seals has been established. Top-10 most similar cases are reported back to allow the police officers at the scene to conduct real-time analysis to identify authenticity, prevent fraud, and collect evidence at the same time.

The system works in three steps. First, it identifies the location of the seal on the document and extracts it, and performs preprocessing such as background removal and correction to remove noise. Then use the pre-trained model combined with SVM to extract global features for quick comparison and preliminary screening. For the relatively similar images screened out, we used SuperGlue to further extract local features and conduct point-to-point comparisons. Finally, the detailed differences of the regional features were reported for police reference. Tests on the police's real database show that the accuracy for identifying counterfeit government seals is over 97%, and the accuracy for forged company seals is over 92%. This system can find out Top-10 fraud cases that the same forged seal has been used recently in a few minutes. It provides the police with immediate identification of authenticity on the spot.

Contact Us for related product information: https://www.hysontech.tw

Low Carbon Smart Electronic Material System for P/N Monitor and Removal

HysonTech team developed an intelligent and low carbon water purification system based on the electrochemical-driven layered double hydroxides material that can monitor the concentrations of nitrogen (N) and phosphorus (P) in water, and remove or degrade 90% of them by electrochemical-driven mechanism in one day. This system can simultaneously monitor and process nitrogen and phosphorus concentrations, helping industries such as industry and aquaculture to solve wastewater discharge pollution problems.