An AI-Powered Predictive Maintenance Platform for Testing AV Components

[AUDIO LOGO]

Hi, I'm Seongil Lee. I'm happy to share with you some of my recent study about an AI-powered predictive maintenance platform for testing autonomous vehicle components at HL Mando. I spent 25 years of experience in the reliability engineering and testing vehicle and its components. This presentation will show you step-by-step guide to producing big data to have an AI-powered maintenance platform service on the web.

I want to give you a little bit a glimpse of what Mando is. HL Mando is a worldwide automotive component supplier. HL Mando was incorporated back in 1962 and has total of 41 manufacturing and R&D locations around the world. We are currently supplying our technologies and products to 60 customers.

And our product could be categorized into four sections-- automotive, such as parts, steering, suspension and brake, and mobility solution and ADAS and software. We focus on developing advanced technologies that form the basis of automotive safety, autonomous driving solution, and vehicle motion control.

Autonomous vehicles add new technology like X-by-wire, which needs high performance and more reliable characteristics on its own. X-by-wire system has made mechanical connection obsolete to have more freedom in packaging and maneuvering. So safety should be guaranteed before producing by increasing reliability testing.

This brought increase in quantity and quality of tests. And large amounts of investment in testing facility was needed in a short time. So we decided to develop our own in-house test benches, which have predictive maintenance capability. With the advancement of Xyy technology, HM Mando has been preparing for Smart Lab.

At the beginning, we focused on the automation and data-driven technology. But today, we have reached more capability to manage and predict the state of future lab by using all these elements like IoT, PHM, and AI. Without Smart Lab, accumulating big data from benches must have been more difficult and inefficient.

How Smart Lab comprises connected test benches, IoT-based monitoring system, and AI-powered big data training and predictive analysis-- to get the right big data from tests and also to reduce cost of development, developing in-house test benches for inevitable. Connected test benches, which is capable of gathering and transmit them to our MPLS server, made all the task flow automated.

You see the advantages of new test benches construction by using in-house actuators and maintenance platform and high performances by using and applying new control algorithm. Building the whole Smart Lab was supported by MATLAB products. Designing a physical model in Simscape helped understand the capacity of new test benches.

And detail control and parameterization were implemented by Simulink. And third, real test and signal processing were conducted by Speedgoat. Lastly, PHM maintenance platform were firstly developed by Web App Server and recently replaced by MPS server.

Building new concept of test benches were interdisciplinary project ranging from communication, cooling, kinematics, and sensor signal processing. In the process of development, control algorithms were patent-protected intellectual property.

Best benefit from new test benches is capability of web-based maintenance platform-- maintenance strategies for conventional test benches for corrective and preventive maintenance, which is very costly and time consuming. But new smart maintenance platform was created to reduce cost and enhance performance.

Also, we applied most new technologies like Edge computing and the AWS DB, Web App Server, and MPS server. Each components of new test benches were modeled and simulated using digital twin of physical model. Metrics Simscape helped build new control algorithms in system level and implement simulation testing.

Two example models show how we create custom component model using MATLAB-based to Simscape languages. Detail features of Simscape helped enumerate the model, leading to reality.

Control architectures as shown in the picture-- host PC sends command data and receive feedback signals from benches and controller. Speedgoat adds a higher level controller, thus real-time target machine, executing 10 kilohertz control task and connect the motor inverters and test samples through various communication, such as etherCat, CAN FD, and analog, and so on.

Lower level controller, which is embedded in test bench, execute task of focal adhesion and function for the safety of the motor. This slide shows how complete three axes steering durability test stand works. Rotate rotary actuator at the center x has a driver. And two other linear actuator x has a tire and suspension of a real car, relating to the driver's action.

New benches with ILC, which means Iterative Learning Control, shows better performance than existing benches. Target RMS error of false control below 10% was attained with a good margin. This is the PHM basic architecture. As shown here, two different paths could compare and differentiate the prediction of the test. They help each other for the precision of the test result.

Data from a real test bench passes to the Edge computer, extracting features and learning with AI and then stored it in a database DD-- and finally, showing real-time CVM and PHM status here. Another data flow from virtual bench goes through different analysis set and working for simulation and tuning.

PHM platform architecture is composed of main EMS system and individual monitoring system. EMS system-- what is called Enterprise Management System-- analyze data in real time and controls the AI-based functions like CBM, PHM, and synthetic monitoring system. It also displays the result of analysis to individual monitoring system in which users can log in and watch the status of equipment under control.

RUL-- Remaining Useful Life-- estimation is the key feature of our PHM system. We analyze data according to its model type. For similarity and degradation models, we use the AI-based analysis. And for survival model, we chose to apply physics of failure because we have lots of data and knowledge in this area.

Collecting live data takes years and even more difficult to get the data in normal conditions with multiple samples. Let's think about it. If we gather components of real car in normal situation, it will take 3 or 10 years. I don't know. But by choosing to collect data in accelerated life testing, which is well correlated to market and mileage, then we have very easy and efficient way for collection of this data.

By implementing accelerated testing with good acceleration vector, life and degradation data were more approachable to data scientists like me. So I want to recommend you to implement this kind of testing to reduce time and cost for the collecting of big data. I want to give you one example of estimating the life by practicing reliability technique, which is physics of failure.

This is what is called fatigue. When geometry, loads, and material information are at hand, then life could be calculated by the following algorithm, which is called Miner's rule. If damage is 1, it means failure occurs. This is only the real example of estimating degradation life of a plastic bearing. We conducted live testing of this sample on our benches by loading and unloading and gathered multiple life curves like this.

So we gathered the data and preprocess the data and condition indicators and train the model. And we can deploy the RUL by using MATLAB Machine Learning Toolbox. PHM platform consists of three basic functions. One is the information app, and second is monitoring app. The third one is tuning app.

Information app deals with basic information like equipment specification, performance, and install and repair history. So we can identify the equipment manual and specific specification and most basic information here. Second is the monitoring app. In this app, real-time monitoring is possible by running an alarm system.

And RUL estimation is also displayed in real time. Third one is the tuning app. Here, you can parameter tuning, and simulation is possible.

My last slide are these key takeaways and conclusion. Our whole process for the big data processing is accomplished by making connected test benches and real-time PHM platform. Newly developed equipment and platform helps to accelerate the maturity of our product, like autonomous components like X-by-wire system. My further plan is to apply this technology to other areas like robots and medical environment. Thank you for listening.

[AUDIO LOGO]