How much fuel does a load-sensing hydraulic system save over a plain open-centre one? The heavy-hydraulics research programme had been answering that with models and invented cycles for two weeks. It was blocked waiting for one thing: a measured trace with pump pressure, actuator pressures and fuel on the same clock.
A Zenodo search turned one up. Linjama, Huova, Tammisto, Tervonen and Ahola at Tampere University published 440 seconds of a 6 t wheel loader loading gravel (CC BY 4.0). It has 2 ms samples of the lift and tilt cylinder pressures and positions, the outlet pressure of the 63 cc variable work pump, engine speed and the ECU fuel rate, from 18 short Y-cycles of about 21 s each. We wrote four predictions in the project log before computing anything, then ran them (measured_loader.py).
What the machine measured
- With no flow demanded, the pump idles at a 30.8 bar standby pressure, so this really is a load-sensing system.
- While the boom or bucket moves, the pump runs a median 24.9 bar above the highest load pressure. That is the first measured margin in the programme. The figures it had used came from valve manuals (14 and 25 bar), and the measured value sits between them.
- The actuators receive 73% of the pump's hydraulic energy.
The model holds
The programme's load-sensing model is one line: pump power is total flow times (load pressure plus a constant margin). With the measured median margin it reproduces the measured pump energy to 2.3%. Across the 18 cycles the error runs 1.3 to 3.4% (95% bootstrap). This is a weak test, because both sides share the same flow, but it shows a constant margin is a fair simplification here.
The open-centre penalty is smaller than the models said
Put a fixed 63 cc open-centre pump on the same motion. It would use about 1.46 times the load-sensing pump energy and add about 12% to the machine's fuel. Across the sweep of unsourced settings that becomes 1.24 to 1.79 times and 6 to 22% more fuel. The programme's first, invented cycle had load-sensing at 47% of open-centre. On the real cycle it is 56 to 81%. Two of the four predictions (open-centre at 1.5 to 3 times, and 10 to 30% more fuel) fail across the sweep.
A bug, and a threshold
The first run gave a much bigger penalty: 2.29 times and 35% more fuel. That was a bug. Filtered flow signals are almost never exactly zero, so the code treated the boom as moving all the time. The simulated open-centre valve never opened its bypass, and the pump paid full flow at the boom's holding pressure for 440 seconds. The fix adds a small-flow threshold below which the bypass is taken as open. We chose it after seeing that run, and it moves the answer more than any valve figure: from 1.7 times at 0.5 L/min to 1.3 times at 5 L/min. A real working cycle spends a lot of time at small flows. The step-shaped cycles of earlier milestones hid this. The next step replaces the threshold with the bypass-closing curve from a published valve manual.
Limits. This is a wheel loader, not an excavator, with one operator and one record. The open-centre machine is a counterfactual that runs the same pump through the same motion. The valve pressure drops at this flow are swept, not sourced. The flows are the dataset authors' estimates from piston velocity. Nothing was run on a machine. Heavy machines are dangerous, and none of this is a design or a safety system. Written with AI assistance (Claude Code); the dataset's authors have not reviewed it.