Engineers at Purdue University are developing a system that will enable people to search huge industry databases by sketching a part from memory, pencilling in modifications to an existing part or selecting a part that has a similar shape.
“It’s like a special kind of Google that lets you search for parts based on their three-dimensional shapes,” said Karthik Ramani, a professor of mechanical engineering and director of the Purdue Research and Education Centre for Information Systems in Engineering, or PRECISE.
Company databases are sometimes so vast that employees are often unable to find a specific part, meaning a new part must be created from scratch.
“You are looking for the proverbial needle in a haystack,” Ramani said. “You have to remember that product variety and complexity have increased drastically.
“Just a single commercial airliner has more than a million unique parts. Such a search method could save millions of dollars annually by making it unnecessary to design parts anew and enabling you to mine for other knowledge, such as past decisions regarding costs and design advice about the part.”
The method will be detailed in a research paper to be presented during the 20th International Conference on Data Engineering in Boston, sponsored by the Institute of Electrical and Electronics Engineers’ Computer Society.
“We take a 3-D model of a part and convert it into a bunch of small cubes called voxels, which stands for volume elements,” Ramani said.
The system uses complex software algorithms to convert the voxels into a simplified “skeletal graph” based on “feature vectors,” or numbers that represent a part’s shape.
“Like our skeleton, it represents the bare bones of a part’s shape and features, such as how many holes it contains and where the holes are located,” Ramani said.
People can select an inventoried part that resembles the desired part and ask the system to find a “cluster” of like parts. Users also can sketch the desired part entirely from memory, or they can choose a part that looks similar from the company’s catalogue and then sketch modifications to that part. The system then assists in finding the desired part.
Not only will the system enable employees to find parts, but it will provide access to valuable background about how the part was produced, including details about machining and casting, which, in turn, provides information about how much it costs to make the part.
“Corporate memory is short,” Ramani said. “People leave, managers come and go. They forget file names and project names. This type of system allows you to retrieve your own company’s knowledge, your own company’s history.
“Let’s say there are 1.3 million parts in your inventory. If you are trying to design a part and you can find something similar that was produced in the past that has a lot of value.”
Design engineers spend about six weeks per year looking for information on parts, he said.
“The shape-search system will allow engineers to cut this time down by as much as 80 percent,” Ramani said.
A series of experiments in which people used the system to find parts showed that it has an accuracy of up to 85 percent.
Findings being discussed during the conference deal with information about how the system indexes parts. The indexing represents parts on various levels of sophistication, ranging from the skeletal graphs to more complex, detailed information.
“It’s very much like how you would index a book,” Ramani said. “You take the geometry of the part, extract its features and then index it.”
The system retrieves a group of parts that resemble the sketch entered into the search.
“Then we allow a person to tell the system which ones are the closest matches,” Ramani said. “And as a result of that feedback, which we call relevance feedback, the system begins to understand a little bit more about what you are looking for and it starts focusing on that.”
Simplifying parts into skeletal graphs is a critical factor that makes the shape-search system possible. The system also allows the user to fine-tune the search to specify aspects such as whether a part was created by casting or machining.
“So the search takes place through multiple representations of the part in a multistep process, which is very important,” Ramani said. “This is not a simple, single-step approach that others have tried.
“We have created the first engineered-part search system. It takes into account the many design and production steps involved in making parts.”
Relevance feedback requires algorithms that use “neural networks,” or software that mimics how the human brain thinks.
“What you want to do is bridge the gap between what’s in your head, your idea of what the part looks like, and what’s in this huge inventory of parts,” Ramani said. “That is not a trivial problem.”
Experiments showed that the accuracy of the multistep search strategy was an average 51 percent higher than a single-step search.
The system is not quite ready for commercialisation.
“We have solved significant problems, but there are remaining challenges,” Ramani said.
The engineers are working with Zygmunt Pizlo, an associate professor of psychological sciences at Purdue, to help improve the system by incorporating information about human perception.
“Working with a psychology expert will help us design experiments aimed at bridging the gap between the human being and the system,” Ramani said. “We are putting a lot of effort into improving the human interface.”
A patent has been filed, and a private company, Imaginestics Inc., in the Purdue Research Park, has agreed to license the technology.