AI to be rapidly adopted in industrial applications
DX through AI is needed as a means to overcome COVID-19
AI model development, one of the advanced system implementation processes


As the government has issued guidelines for 'social distancing' to prevent the spread of COVID-19, MathWorks has decided to hold the MATLAB Expo 2020, which is held annually, online for a total of three days (July 2, 9, and 16).

Prior to this, MathWorks held a media briefing via Zoom on the 2nd and announced the five major AI trends for 2020. Jim Tung, head of business and technology strategy and analysis, who made the announcement, predicted that “AI will be rapidly introduced into industrial application areas.”

According to Gartner, the average number of AI projects for companies is expected to grow by about tenfold from 4 in 2019 to 35 in 2022, and ‘integrating AI into systems’ is becoming a top priority for companies. Meanwhile, 'low AI technology proficiency (56%)' and 'data quality issues (34%)' were pointed out as obstacles to successful AI adoption, so solutions that can resolve these are expected to attract attention.

In addition, while some companies have scaled back their operations due to the global pandemic caused by COVID-19, some companies and organizations are choosing digital transformation as a means of overcoming the crisis. Professor Eun-ok Jeong's research team from the Department of Mathematics at Konkuk University developed a Korean-style COVID-19 spread model using MATLAB and obtained research results that can serve as scientific grounds for the government's infectious disease spread response policy.

Professor Jeong’s team developed an SEIR model that takes into account not only the number of confirmed cases but also various variables such as changes in people’s behavior, social distancing, and the implementation of the delayed school opening system through a series of studies utilizing MATLAB’s mathematical modeling technology. Through this, it was revealed that the government’s COVID-19 control and prevention policy is effective in preventing additional confirmed cases.

Tung, the head of AI, listed the following as the five major AI trends for 2020, taking into account COVID-19: ▲easing of AI technology barriers and improvement of data quality ▲increased design complexity of AI-based systems ▲easier deployment of AI to low-power, low-cost embedded devices ▲reinforcement learning expanding from games to real-world industrial applications ▲AI models developed based on a large amount of AI model training data generated through simulation.

◇ Alleviating AI technology barriers and improving data quality

As AI proliferates, the number of engineers and scientists involved in AI projects is increasing. Today, they have easy access to deep learning models and research results from the community, and they are receiving great help from the early stages of AI projects. Additionally, while past AI models were based on image data, they can now process a variety of sensor data, such as time series data, text, and radar.

“MathWorks enables engineers and scientists to apply their domain expertise to building AI-based products and services without having to learn programming,” said Tung. “We provide algorithms that support classification and prediction, a variety of pre-built models from the community, and hundreds of examples that help build AI models across a wide range of engineering applications.”

◇ Increased design complexity of AI-based systems

As AI becomes more capable of handling more types of sensor data through training, engineers are applying it to a wide range of systems, including self-driving cars, aircraft engines, industrial plants, wind turbines, and cloud-based streaming systems.

In complex systems, AI models have a significant impact on overall system-level performance. Therefore, the development of AI models today should be viewed as one step in the process of effectively operating complex systems.

Designers should consider adopting “model-based design” tools to simplify the design of complex AI-based systems: simulation to help designers understand how the AI interacts with other parts of the system; integration to help designers try out different design concepts within the context of the complete system; and continuous testing to help designers quickly identify weaknesses in AI training datasets or design flaws in other components.

“MathWorks helps verify the effectiveness of AI models through simulations before deploying them to hardware,” said Tung. “In particular, Simulink helps perform tests through simulations and reflect the results to quickly perform iterative design processes.”

▲ [Image = Mathworks]

As an example, Tung cited the case of Voyage, which developed a level 3 autonomous vehicle in less than three months. Voyage was able to accelerate the process from ideation to road testing using a unified model, and tune the controller based on experimental data to achieve its goal.

◇ AI deployment to low-power, low-cost embedded devices made easy

Until now, AI has used 32-bit floating-point operations, such as those used in high-performance computing (HPC) systems such as GPUs, clusters, and data centers. It could not be implemented in low-power devices that use fixed-point operations. Advances in software tools now allow us to provide AI inference models that support various levels of fixed-point operations.

The ability to deploy AI on low-cost, low-power devices has opened up a world of design possibilities for engineers. “We can now apply AI to electronic control units (ECUs) in vehicles and other industrial embedded applications,” Tung said.

In doing so, he said, “MathWorks supports deploying AI models developed with MATLAB or Simulink to all environments, including embedded devices, enterprise systems, edge, cloud, and desktops, without rewriting, through a code generation framework.” He also added that it supports an automatic code generation function for model deployment, which can eliminate coding errors.

◇ Reinforcement learning extends from games to real-world industrial applications

This year, reinforcement learning is expected to expand beyond gaming into real-world industrial applications such as autonomous driving, autonomous systems, control design, and robotics. And it is expected to contribute to improving large-scale systems by training virtual models that reflect conditions that are difficult to implement in the real world.

To make this possible, technologies are needed that support the following: △easy reinforcement learning policy construction and training, △generation of large amounts of training simulation data, △easy integration of system simulation tools with reinforcement learning agents, and △code generation for deployment to embedded hardware.

“Reinforcement learning agents can improve and optimize performance by increasing speed, minimizing fuel consumption, and minimizing response time,” said Tung, citing the example of improving the driving performance of autonomous driving systems. “Reinforcement learning models can also be integrated into fully autonomous driving system models that include vehicle dynamics models, environment models, camera sensor models, and image processing algorithms.”

MathWorks' Reinforcement Learning Toolbox provides built-in, customizable reinforcement learning agents and supports modeling reinforcement learning environments in MATLAB or Simulink.

In addition, it supports reference examples such as △Deep Learning Toolbox that supports reinforcement learning policy design, △Training acceleration function on GPU and cloud, △Training policy verification function through simulation, △Deployment function to embedded devices, and △Reinforcement learning for autonomous vehicles.>
“Working with Microsoft, MathWorks will enable customers developing applications in robotics, energy, manufacturing, and process optimization to streamline their workflows on the Microsoft Autonomous Systems platform,” said Tung. “This will enable them to scale and perform simulations on MATLAB and Simulink models in Azure.”

◇ AI model developed based on a large amount of AI model training data generated through simulation

Poor data quality is the biggest obstacle to successful AI adoption. MathWorks predicts that simulation will help solve these data quality issues easily in 2020.

Improving the accuracy of AI models through training requires a large amount of data. However, unlike data related to normal system operation, data on truly necessary abnormal conditions or serious failure conditions are difficult to obtain.

“For a health predictive maintenance application that can accurately predict the remaining life of a pump in an industrial setting, it is difficult to obtain failure data because failures in physical equipment are rare,” said Tung.

The best way to obtain large amounts of high-quality data for training AI models is to generate data by running simulations that mimic dysfunctions and then train the AI models using the synthetic data to develop accurate models.

By simulating the model of the aforementioned pump, signals indicating failures can be generated, and AI models can be trained based on these signal data and deployed to actual systems in industrial settings to predict future failures.

Simulations can play a key role in building AI algorithms that respond to rare events in industrial settings. “MATLAB and Simulink help engineers generate and integrate training data for infrequent events through simulation at appropriate intervals,” said Tung.

Finally, Mr. Tung said, “Many organizations are currently focusing on developing AI algorithms,” and advised, “In order to successfully bring AI-based products or services to market, rather than simply AI algorithms or models, AI must be integrated into the overall system design workflow.”