The heavy-hydraulics programme asks whether a load-sensing pump is worth its price over a fixed open-centre pump on a small earthmover. Its first answers came from a force model of an excavator's dig cycle. Later answers came from one measured machine, a 6 t wheel loader. One gap stayed open throughout: we had never seen pressure data from a real excavator. Now we have.
Where the data came from
Searches of Zenodo, Figshare and Mendeley Data turned up no open excavator pressure trace. A Hugging Face dataset, fuxi-robot/excavator-motion (CC BY-NC-SA 4.0), is published as a motion dataset for autonomous truck loading. One of its three machines, a large excavator in a folder named 490, also logs eight hydraulic pressure channels at about 10 Hz. They sit beside the boom, arm, bucket and swing joint angles.
The files are 4 to 5 GB each because they carry camera images. The arrays we need are a few MB, so the script reads only those, using HTTP range requests: about 30 MB of each file over the network instead of 5 GB. Nothing from the dataset is committed to the repository. The script downloads what it needs and prints the derived numbers.
What it shows
We wrote five predictions into the research log before writing the analysis. Then we ran it on all ten recordings, 76.7 minutes in total.
- One channel is dead in every file, flat at -60.6 bar. The other seven are live.
- While the boom, arm or bucket is moving (87% of the time), the highest live channel has a median of 174 bar. The 95% bootstrap interval over files is 167 to 180 bar.
- That is 5.3 times the 33 bar of our force model, and 1.3 times the wheel loader's 134 bar.
- The figure moves by less than 4 bar if we change the motion threshold, change how pressure rows are paired with angle rows, or drop the motion filter altogether.
The channels carry no labels. The highest live channel is therefore an upper reading of the highest load pressure, since it could be a pump or return-side pressure. A 49 t machine also runs a higher relief pressure than the small machine our model describes. To allow for both, we also report pressure as a share of the machine's own top pressure. By that measure the machine works at 0.39 to 0.43 of its top pressure. The force model works at 0.18, so the measured share is still more than twice the model's. On a small machine with a 180 bar relief, that share means about 71 to 78 bar, not 33.
Why it matters for the circuit choice
Earlier steps showed that the break-even margin, the pressure margin a load-sensing pump can carry and still beat open-centre, is roughly (volume ratio minus one) times the working pressure. Swapping in the measured pressure at the model's volume ratio moves that break-even from about 11 bar to between 24 and 59 bar. Commercial load-sensing pumps run margins of 14 to 25 bar. So the synthetic results had understated the case for load-sensing. On measured excavator pressure, load-sensing wins unless the excavator's flow is far smoother than our synthetic cycle.
Two of the five predictions failed. We guessed that one channel would behave like the pump pressure: no single channel dominates. We also guessed the excavator would land within 30% of the loader: it came in at exactly 1.30 times, just outside the band.
What is still missing
These files contain no flow data. The other half of the break-even, how peaky an excavator's flow demand is, is still untested on a real machine. The same dataset has a smaller, 7.5 t class machine with joint angles only. Combined with published cylinder geometry, those angles could give that flow. That is the next step.
The script is experiments/measured_excavator.py and its output is measured_excavator_results.txt. Section 3.7 of the paper draft has the detail.
This is analysis of public data, not a machine design and not safety advice. Written with AI assistance (Claude Code).