This series of papers is part of Google’s Open Random Vibration Testing of Off the Shelf Data Center Hardware Project, available and distributed through the project’s Github repository.
This paper follows Part 1 (Looking inside the black box: How Google Uses Real-World Vibration Data to Improve Data Center Hardware Testing), which outlined how Google captured and analyzed random vibration data from Google’s global supply chain, and explained why it is important to capture accurate data using thoughtful measurements and analysis techniques.)
In this post, we'll look at the following:
Over-testing vs. Under-testing for Real-World Data vs. ASTM Standards
Figure 1.
Engineers can and do spend decades learning how to capture and analyze random vibration data properly. At the end of the day, the results should be easy to understand and impactful in day-to-day work, such as;
We will continue to work toward that in these papers. In Part 1, we described the measurement and analysis methodologies in detail. Part 2 will focus on their application on real-world data and key takeaways from Google’s US measurements.
Sine Vibration, 1G zero to peak, 10 Hz, 360 seconds
Figure 2.
Last time, we discussed how traditional PSD plots remove valuable information from real-world time history data. To put it in simple terms, when we look at sine vibration, two signals are considered equivalent if:
These basic metrics enable us to extract meaningful insights about the actual transportation environments, which was previously difficult with traditional methods. Here are some key takeaways from Google’s real-world US Data:
|
Key Takeaways |
Significance |
|
1. ASTM d4169-14, field data vs. shaker table profile |
Very similar |
|
2. Air ride suspension vs. steel leaf suspension |
Significant difference |
|
3. Comparing air ride suspension in different region of the US |
Significant difference |
|
4. Comparing locations within the same vehicle with air ride suspension |
Notable difference |
|
5. Visualizing cycles of vibration over time |
Significant difference |
|
6. Visualizing “Over-testing” and “Under-testing” |
Unavailable from past method |
Table 1.
Field Data (Steel Leaf) vs. ASTM 2014 (Steel Leaf)
Figure 3.
ASTM’s d4169-14 truck profile, level 2, generally represents real-world data really well. Factors that contribute toward such alignment might include shorter transit duration, smaller vehicle sizes, and most importantly, consistent road conditions. In our experiences, vehicles with such profile are primarily used for transits within local areas (15 to 30 mins), which greatly eliminates variability in road conditions. Such consistency is not observed when examining data from vehicles represented by ASTM’s d4169-16 truck profile (air ride).
Field Data (Air Ride) vs. ASTM 2014 (Steel Leaf)
Figure 4.1
As shown in 2.1, real-world data of vehicles with air ride suspension looks very different from vehicles with steel leaf suspension (represented here with ASTM d14169-14 truck profile). The locations of high cycle are very different, along with locations of high G amplitude. The ASTM truck profile was changed for 2016 with good reason. Still, it is difficult for shaker table profiles to accurately replicate the overall shape seen in real-world data, with correct cycle counts for specific frequencies.
Figure 4.2
We captured shaker table data of the 2016 profile during one of our lab experiments, and saw remarkable differences with the 2014 profile as well as real-world air ride data.
Google US Midwest 1 (Air Ride) vs. Google US Midwest 2 (Air Ride)
Figure 5.
Another way to look at the data is by creating heatmaps like figure 5. Cycle count is represented with color mapping, and the two dimensional view allows us to focus our attention on the relationship between frequency (2 to 300 Hz) and amplitude (-2.0 to 2.0 G).
Such visualization really allows us to appreciate the difference between two sets of data. In this example, the two data sets are supposed to be the same because they are both from vehicles with air ride suspension. But as we can see, they look very different because the vehicles traveled in different regions within the US Midwest. Feedback from people who live in Midwest 1 says road conditions there are much worse, which explains the higher amplitudes at low frequencies. Worse roads make the vehicle bounce more, which leads to more vibration.
Google US Midwest 2 (Front) vs. Google US Midwest 2 (Rear)
Figure 6.
Another interesting finding came from the same vehicle traveling in Google US Midwest 2. Sensors were placed near the front of the trailer (left figure) and the rear of the trailer (right figure), and their data distributions were compared and shown in figure 6. There are notable differences between the two plots - 1.) cycle counts for low frequencies (<50 Hz) are noticeably higher for the rear of the trailer, and 2.) amplitude for high frequencies (>90 Hz) are significantly higher for the front of the trailer.
One explanation might be that the rear of the trailer consists of a long, heavy load platform on soft suspension and long wheelbase, which leads to large vertical displacements at low frequencies, while the front is picking up high frequency vibration from combustion pulses, gear-mesh tones, turbo whine, and universal-joint imbalance from the tractor unit.
Figure 7.
Having precise timing of peak amplitudes also allows us to visualize the distribution of cycles over time. This can offer clues for road conditions if the transit route is known and timing correlated with it.
For example, the US Midwest 1 data set was broken down into 15 minutes segments, each showing a distribution of cycle count vs. amplitude. Figure 10 shows data for 9 to 11 Hz, and shows higher cycle counts in the first half of the data set.
Did the vehicle travel faster during the first half of the shipment? Were the road conditions worse? Did the vehicle paused more during the second half of the shipment? These are all interesting questions to explore as we look further inside the data.
Such visualization can be made for different frequency bands to see if it makes any difference, and there appears to be when we look at several other frequency bands between 7 and 21 Hz. The distribution of cycles over time are not all uniformed, which makes these plots really interesting to look at.
Visualizing cycles of vibration over time, US Midwest 1, 7 to 21 Hz
Figure 8.
3D histograms allow us to compare real-world data with shaker table profiles, and visually identify areas of “Over-testing” (cycles of shaker table vibration that exceed real-world data) and “Under-testing” (cycles that fall short of real-world data). For example, when we take the difference between US Midwest 1 and ASTM d4169-14 (scaled to 2 hours), we generate the following 3D histogram, which shows area of “Over-testing” in red and “Under-testing” in blue:
Visualizing “Over-testing” and “Under-testing” between US Midwest 1 and
ASTM d4169-14 Truck Profile
Figure 9.
Alternatively, this can be examined in a spreadsheet, where the exact difference of cycle count is calculated and shown in each cell. Orange/red colors representing “Over-testing”, blue/purple representing “Under-testing”.
Visualizing “Over-testing” and “Under-testing” in Spreadsheet
Figure 10.
Visualizing “Over-testing” and “Under-testing” in Spreadsheet
Figure 11.
As we can see, in this comparison, there are 100,000+ cycles of “Over-testing” at peak amplitudes around 0.1G to 0.4G, between 10 Hz and 22 Hz. Depending on the product, how it is packaged, and the specific failure mode, it may or may not make a difference.
At the end of the day, efforts put into this are meant to help us make more informed decisions in product design, product reliability, packaging design, and design verification/validation. The measurements techniques and test methodologies will continue to be debated and improved, as all engineering knowledge does, but new thinking provides new answers that previously were not easy or possible to obtain.
Q1: Are road conditions and vehicle conditions the same everywhere around the world?
A: No, they are not.
Q2: Can measured data help us evaluate how appropriate test standards are for our unique situations?
A: Yes, they can.
Q3: Do we have a better sense of the range and variety of random vibration my products are expected to see?
A: We have a much better sense now than we ever did previously, but the world is big and the supply chain is incredibly complex, so we need more data coverage to be certain.
Q4: Will we ever get to a point where all we have to do is enter the addresses of the origin and destination of a shipment, and get a customized analysis/test plan for our specific route of transportation?
A: We will, with multiple solutions that are independently developed by people passionate about this area of work.
The methods and techniques developed in the course of this project are meant to be thorough so that it can independently be replicated and evaluated by anyone interested. But it is still important to remember the result should be simple to understand and the learnings easy to implement. I will continue to work on that as the project progresses.
Now that a lot of the technical details are published, it’s nice to be able to switch gears and release results/findings for a change because that’s what matters to most people. You will see that more and more as we move forward. It’s hard to look up and see what’s really happening around us, when we’ve spent so much time trying to look inside the details. A.I. data centers, LLMs that write codes and conduct research, self driving cars, robots that backflip.
I hope the material released thus far is enough to spark your interest into thinking about all of this from a different perspective. It’s been more than 7 years since I began measuring data in the real world. Even as I look at the data today, I still think to myself, “this stuff is still very interesting.” Can shock and vibration keep up with time and technological advancement? I think it can, if we continue to challenge ourselves.
As a company next door right in the heart of Silicon Valley once famously advertised: Think Different.