Honda engineers design to safety and fuel efficiency regulations using AI
At Honda R&D, we’re infusing AI into our vehicle body design process with our Expert Knowledge System, a solution to make all of our engineers as good as our best.
“Our goal is to increase our design accuracy to reduce late stage simulations by as much as 80 percent. Using better information earlier in the design process will lead to fewer failures in physical tests.” – Shigeto Yasuhara, Chief Engineer, Honda R&D Co., Ltd., Japan
Our department designs for crash safety of passengers and drivers. Our work requires intense physics and rigorous physical analysis and simulation. We wanted to apply AI to help our engineers find the best design solution faster — ultimately contributing to safer and more eco-friendly vehicles.
Making invisible expertise available
Between technological advances, radical transformation in the economy, and environmental and safety regulations, twenty-first century automakers face unprecedented challenges.
Connected, autonomous vehicles, autonomous driving, sharing economy, electric vehicles, environmental regulations for zero emissions vehicles and safety requirements intended to protect drivers and passengers are becoming increasingly difficult to meet. Engineers who design and develop to these standards — and who must balance intricate tradeoffs of production and esthetics without compromising safety or quality — need to be experts.
Our organization is only as good as the intellectual capital we accumulate over time — the combined intellect, imagination, information and experience of our engineers. We wanted to multiply this knowledge and expertise across large teams of engineers over time and along each engineer’s career progression, across all Honda models, product lines, countries and languages. We have the creativity, intellect and information, and now we’re on a path to join those together so that our engineers can rapidly access and understand established knowledge and ignore design paths that have been invalidated.
We want our design engineers to benefit from our collective expertise to spark new ideas in directions that show the most promise. To accomplish this goal, we have moved beyond merely collecting the structured and unstructured data contained in digital systems, moving to capture the data locked in the minds of our expert engineers — what we call invisible unstructured data — our deepest, richest data, the collective knowledge of our people.
Infusing intellectual capital into the design process
When creating new models or features for existing products, Honda engineers must think about many factors, from safety to regulations, productivity to costs. And we must maintain the same design methodology across hundreds of engineers in dozens of locations around the world.
But not all engineers work in the same way, bring the same experience to the job, or experience technology in the same way. We found that less experienced engineers spent more time running design simulations, whereas highly experienced engineers ran fewer. The more experienced, expert engineers drew upon their own — and established institutional knowledge — to rule out inferior options more quickly, allowing them to focus solely on potential new options. Their ability to work in this way is a product of years of experience, knowledge they absorbed and could access more intuitively.
When we realized this was happening, we tried to equip less experienced engineers with information — lots of information — from images, graphs, drawings and photos. But this method was static, difficult to access, and isolated design activity within the single vehicle component, prohibiting the engineer from designing products holistically.
This inevitably led to more design simulations and a greater failure rate in later stages of the design process.
Using AI to collect knowledge
We understood the fundamental problem: the knowledge the senior engineers possessed had to be made accessible to their juniors. We contacted IBM to help us apply AI to solve this challenge.
We learned that our experts don’t approach problem solving in a linear fashion. Instead, they have an interconnected network of ideas. Some ideas arrive with more associated ideas or related ideas to be considered. They can, using their tacit knowledge, trace which one among all alternatives has their highest confidence level. Now we use AI to align to and augment their mental models.
To capture and make available the accumulated intellectual knowledge of our engineering organization, we defined several layers of knowledge associated to different design stages. We were able to capture language and concepts derived from engineers directly, what we call invisible unstructured information, the priceless knowledge historically locked away in engineers’ minds. Using Watson Speech to Text and a patented annotation language, we mapped this information to build our knowledge models graph.
With Natural Language Processing our expert engineers can express themselves naturally, even using complex equations and explanations about how they approached a problem. The system enables us to capture information about the parts of the problem they accepted as given, their design constraints, and the network of ideas they previously explored. Using the new system, we expect to see a 50 to 80 percent reduction in the time required for knowledge modeling compared to a typical manual modeling tool.
“We learned that our experts don’t approach problem solving in a linear fashion. Instead, they have an interconnected network of ideas. Some ideas arrive with more associated ideas or related ideas to be considered. They can, using their tacit knowledge, trace which one among all alternatives has their highest confidence level. Now we use AI to align to and augment their mental models.” – Shigeto Yasuhara, Chief Engineer, Honda R&D Co., Ltd., Japan
Using AI to design safe and more eco-friendly vehicles
The knowledge models graph aligns closely to the engineer’s thought process and establishes the entities and relationships within the collection and provides data visualization showing the network of ideas, domain ontology, and the respective confidence level of each node based on the expert’s input.
We developed a proprietary language to notate the expert engineer’s thought processes and approach to problem solving. We have begun long-term work to link and validate supporting evidence to each node in the knowledge graph integrating this valuable, previously invisible unstructured information, with digital structured and unstructured data.
The system helps engineers develop new features contextually with larger systems. For example, if an engineer is working on a design for the panoramic glass roof for the Honda Pilot Elite, the design possibilities, and dependencies, can’t be determined in isolation from related components. The engineer must know how to design to meet the specifications of each component and understand the relationships between components.
To determine if a particular feature is viable, the engineer must consider drivability by correctly designing body stiffness, performance in crash tests, roof thickness and other features.
We have tested our Expert Knowledge System to bridge from the spoken word, in each engineer’s native language, to the network of ideas they want to explore. Think of how many engineers, regardless of experience, would express a design problem and the given constraints. The expression is typically a linear expression such as: “I need to reduce the weight of the front bumper face by .005 kg without transferring loads through structural members at 5mph.”
Natural language statements return the appropriate responses by mining the collected expert knowledge. The engineer can see the approach returned by the system, with the rationale and supporting evidence, and know the response is sound.
Drawing on the knowledge contained in the Expert Knowledge System, we had enough information in place to test whether we could ask the system a design question and have the solution search through the information, pursue the right pathways of the knowledge graph, and retrieve the correct information. This important development allows engineers at any level of expertise to ask questions and retrieve the best answers.
Our goal is to increase our design accuracy to reduce late stage simulations by as much as 80 percent. Using better information earlier in the design process will lead to fewer failures in physical tests.