Caterpillar delivers a product line that might be the most diverse in the world. Its equipment ranges from simple, small 13-hp (9.7-kW) excavators to machines that require over 20 railroad cars to deliver, according to Randall Huber, Advanced Virtual Product Development Manager at Caterpillar. The same equipment may need to work high at thousands of feet of elevation to far below sea level, in the arctic or desert. In addition to those challenges are increasing regulation, a dynamic market, and fierce competition.
These challenges are driving a push to the virtual, guided by Huber. “We define virtual product development, or VPD, as physics-based simulation,” stressed Huber. While useful in validation, Caterpillar is now using VPD to develop early concepts and influence design before it becomes fixed for manufacturing. “It is all about driving early discovery in the development process,” he said. While accounting for the typical functions a virtual world can simulate, Caterpillar’s VPD is unique in that it starts with work site productivity. Some of this is done using proprietary software Caterpillar created. Durability, reliability, performance, quality, and efficiency are all qualities defined through using virtual tools. From manufacturing of components to work site deployment, Caterpillar desires to simulate it all.
The ultimate goal—seamless simulation
Like other organizations, Caterpillar is continuously investing in its VPD. It is in the process of creating higher levels of capability through improving both product data management and better physical models. At its simplest, its CAE simulation provides directional guidance, confirming that, for example, Design A is better than Design B without being able to quantify exactly how much or attempting to correlate simulated results with test data. The next level is producing answers that are quantitatively correct—correlated with test data. The next level of capability is to do this rapidly, through standard work procedures and shared models. “Once we can do it fast, then we can explore designs quickly, creating robust designs that will work well and are consistently manufacturable,” he said.
The computing and data management infrastructure for this ultimate, seamless VPD requires plenty of computing horsepower, which is becoming faster and cheaper. It also requires accurate loads prediction—model inputs that accurately capture what the customer will do with the machine based on practical product application knowledge.
Finally, it requires quantifying uncertainty. “We have got to get away from discreet event modeling,” he explained. “The world is way too complex. As we do model simulation, we cannot look at it just as a discrete event.”
Fundamentals and uncertainty
While correlation from model to physical test is vital, it has its problems. How many tests accurately predict the behavior limits of a machine in operation? If a simulation model uses only nominal dimensions and loads, what variations on those will capture real-world performance?
“For example, if we run a real bulldozer over a ditch four times to calculate an average life, and turn that data over to a simulation expert, our analysts have determined it is impossible to expect a single simulation to match the average at all locations simultaneously,” he explained. How does a design engineer interpret such a result? Is it wrong, or was the test data not representative?
To make CAE simulation more useful, Huber and his team turned the tables on the problem. Instead of assuming that the simulation was either correct or incorrect, they ran a series of simulations at varying conditions similar to the test and then compared the resulting distribution to determine if these are statistically from the same data set. In effect, they treat the simulation as if it was run number 5 through 25 of the physical test in his bulldozer example. “When you do that, all of a sudden you get a different thought process on correlation,” he said. Now simulation is used to determine sensitivity to inputs, and as long as the test data and simulation data are determined to be from the same population, they become an integrated set of data. “You can now run a thousand point DoE where you cannot tell which is simulated and which is test,” he said. The next step in their development is to do much of that DoE work up front in simulation, prior to physical testing.
“So as we look at it, our simulation's equations are a function of credible physics and technical competence,” he explained. “But [also] a combination of reasonably correct physics combined with really good statistics.”