Applying AI to business requires an organization's AI capabilities and
Need to improve data quality and eliminate silos across departments
MathWorks Provides Integrated Solutions for Building AI Systems


As digitalization accelerates, many companies are rushing to introduce AI into their businesses. According to Gartner, the average number of AI projects for companies is expected to increase from four in 2019 to 35 in 2022.

However, the professional AI knowledge and experience of domain experts who must design and implement AI systems are still lacking. The success of AI adoption depends on the organization’s AI capabilities, the quality of data, and the elimination of silos between departments responsible for different components of the overall AI system.

On the 18th, Manager Song Wan-bin, an application engineer at MathWorks, discussed the challenges companies face when developing AI systems through a YouTube live webinar. We also introduced MathWorks' unique solutions that support this and shared success stories from various industrial fields.


Step-by-step AI system design process
Manager Song Wan-bin emphasized that in order to design an AI-based system, engineers must consider the following stages: ▲Data Preparation ▲AI Modeling ▲Simulation and Test ▲Final Deployment.

The data preparation stage involves data cleansing and preprocessing. It also includes data augmentation and fake data generation when the amount of data is insufficient. MathWorks provides high-level functions and domain-specific tools to help domain experts with limited AI knowledge and experience carry out AI projects, helping build improved quality data sets.

Domain-specific tools provide interactive UI-based labeling tools optimized for a variety of data types, including images, videos, audio, and signal data, to automate the 'labeling' task of assigning correct answers to data.

The AI modeling stage includes the task of designing and adjusting the model by determining the network structure and learning algorithm type according to the data type and AI introduction purpose. MathWorks supports interoperability with deep learning frameworks such as Python, allowing the completed work of the framework to be imported into MATLAB and used.

The simulation and testing phase includes integration of AI models with existing systems, simulation of AI systems, and testing for verification and validation. MathWorks provides a single platform for AI modeling and simulation, reducing integration risks at the system level and supporting interactive tools that make it easy to train and experiment with machine learning models.

The final deployment stage refers to deploying the AI model or system to an embedded target, enterprise system, or cloud. MathWorks supports both code generation optimized for the deployment target and lightweight AI models for deployment to the edge and cloud.

“Building an AI system is not a one-time task,” said Manager Song. “AI models and systems must be improved by repeatedly performing data preparation, modeling, simulation, testing, and deployment stages.” “To achieve this, we need to eliminate silos between each step, and MathWorks has solved the silo problem by supporting the above process in an integrated environment,” he said.


ML and DL adoption cases of famous overseas companies
Manager Song mentioned that MathWorks’ machine learning solutions have been utilized in tire wear detection by tire manufacturer Bridgestone, customer churn prediction by digital IT product and solution provider Cognizant, energy usage analysis and optimization by building energy management solution provider BuildingIQ, oversteering detection by automobile manufacturer BMW, and fleet data analysis by automobile manufacturer Honda.

He added that there are many successful cases of object detection for not only images, signals, and voices but also 3D data, such as equipment identification by oil company Shell, pathology analysis by biotechnology company Genentech, defective equipment detection by aircraft manufacturer Airbus, and automation and abnormal part detection by auto parts manufacturer Musashi.


Case studies of domestic companies and institutions using MathWorks solutions
Many success stories are being built domestically in line with global trends. Manager Song introduced the cases of Hyundai Steel, Korea Institute of Construction Technology (KICT), Korea Institute of Ocean Science and Technology (KIOST), and its affiliated Ship and Ocean Plant Research Institute.

◇ Hyundai Steel applies pixel-based deep learning technology to steel raw material image analysis

Before introducing MathWorks solutions, Hyundai Steel had analysis engineers manually analyze steel raw material images using optical microscopes. This process took a very long time, and the results varied from analysis engineer to analysis engineer. Therefore, they decided to utilize MathWorks solutions for quantification and automation of steel raw material image analysis. Hyundai Steel decided to apply semantic segmentation techniques among deep learning techniques.

We created an automated labeling algorithm using unsupervised machine learning in MATLAB, imported it into Labelor, and improved the accuracy of the deep learning model by up to 85% by repeating the cycle of labeling, inspection, learning, and reflecting the learning results in Labelor. In addition, the labeling task time, which takes up a large portion of the semantic segmentation task time, was significantly reduced through the labeling automation algorithm.

◇ KICT Develops Deep Learning Model to Support Measurement of Sewerage Wastewater Discharge

In the past, KICT measured the amount of wastewater in sewer pipes using vision and sensor-based solutions, but accurate measurements were not possible due to instability caused by impurities, corrosion, and turbidity.

Here, we developed a deep learning model that precisely measures the height of wastewater in sewer pipes based on image pixels by applying the semantic segmentation technique with MATLAB. The height of wastewater was measured through mathematical calculations using the deep learning model, and the water level detection rate was increased from the previous 53% to 100%.

◇ KIOST Develops LSTM Deep Learning Model for Early Prediction of Red Tide Occurrence

To predict red tide occurrence early, KIOST introduced the LSTM (Long Short-Term Memory) deep learning model, which is widely used to predict time series data other than images.

By preprocessing sea surface temperature and satellite-derived photosynthetic effective radiation sequence data and training an LSTM network with the data, we were able to develop an application with an accuracy of 98% to 100%.

◇ Ship and Ocean Research Institute develops drone-based survivor search technology using deep learning object detection algorithm

The Korea Research Institute of Ships and Ocean Engineering has developed a marine disaster detection model using deep learning object detection. Previously, survivors were searched and rescued visually in the ocean, but an AI-based real-time survivor search solution was developed for efficient and accurate detection.

The AI-based real-time survivor search solution using the YOLO object detection algorithm detects objects and survivors based on drone video data received via deep learning models and UDP communication. The solution achieved an average accuracy of 99% and can replace on-site search.

“MathWorks provides more than 100 examples to support companies’ AI project planning and PoC stages to accelerate the adoption of AI in various industries,” said Manager Song. “Engineers can use the examples as a starting point to customize and apply data and network architecture to concretize AI models and shorten development time.”

He also added that you can get help from MathWorks Korea engineers on MATLAB Github.