A fixed pump pushes 2.4 times the oil a measured excavator uses

Joint motion from a public 7.5 t excavator dataset, turned into cylinder flow: a fixed pump sized for peak demand delivers about 2.4 times the oil used.

The heavy-hydraulics programme asks which hydraulic circuit a small shop could build to get a compact excavator closest to stock fuel use. A rule it found this week says when a load-sensing pump stops beating a plain fixed pump with an open-centre valve. The break-even load-sensing margin is about (r − 1) × p. Here r is the volume a fixed pump delivers while the machine moves, divided by the volume the cylinders actually use, and p is the pressure the surplus is pushed at.

This morning's run measured p on a real excavator: a median of 174 bar while it moves, about five times what our force model assumed. That machine's files have pressure but no flow, so r on an excavator was still a guess. On our synthetic cycles r was 1.43, and on a measured wheel loader it was 1.79.

Getting flow from joint angles

The same public dataset (fuxi-robot/excavator-motion, CC BY-NC-SA 4.0) has a second folder: ten truck-loading recordings, 16 minutes in all, of what the file names suggest is a 7.5 t class XCMG excavator. It records joint angles and joint speeds and nothing else. A cylinder's flow is its piston area times its rod speed. XCMG's XE80E brochure lists the cylinders (boom 110 × 75 × 922 mm, arm 95 × 60 × 885, bucket 85 × 55 × 742), and a dealer listing for the XE75U gives the same sizes to the millimetre. We set each cylinder's lever to its stroke divided by the joint's observed travel, and charged swing at a share of pump flow.

We wrote five predictions down before reading the data. Three were wrong:

  • Which way each cylinder extends matters a lot. Filling the full bore takes about twice the oil of filling the rod side. Across the eight ways to assign directions, r moved by 72%, where we had predicted under 15%. The dataset's own visualisation script settled the boom and the arm: its forward kinematics show a larger boom angle lowers the boom and a larger arm angle pulls the arm in. With those two fixed, the bucket's direction moves r by under 1%.
  • Peak demand ran past the pump. With swing counted, the 99th percentile of demand came out at 240 L/min, above the machine's 219 L/min pump. Without swing, the cylinders alone came to 218.8 L/min. That is the pump's published maximum, matched to 0.1% with nothing fitted, which is what you'd expect from a machine whose top speed is set by its pump.
  • r was higher than we guessed. We registered 1.43 to 2.5.

What came out

The margin column puts each r into the rule at one pressure, 130 bar. That is the 49 t machine's working share of its top pressure, applied to the 7.5 t machine's 316 bar maximum. The 0.78 covariance factor comes from the loader.

Volume ratio rBreak-even margin by the rule, at 130 bar
Synthetic excavator cycles1.4344 bar
Measured wheel loader1.7980 bar
Measured excavator, p99-sized fixed pump2.38 (2.34 to 2.42 over files)139 bar
Same, over 384 lever, sign and swing cases1.95 to 3.0896 bar at the lowest

The excavator moves in 97.5% of samples, but its demand is peaky: a median of about 90 L/min against a 99th percentile of 240 L/min. A fixed pump big enough for the peaks spends most of the cycle pushing oil the work doesn't need, against whatever pressure the heaviest function is holding. Load-sensing pumps are sold with margins of 14 to 25 bar. On this data the margin at which load-sensing would stop winning is 70 to 140 bar, so it wins with plenty to spare.

What this does not show

The flow comes from a 7.5 t machine and the pressure from a 49 t machine, so the two halves of the rule are not from one excavator. Constant levers are a rough stand-in for the real linkage, and the bucket's four-bar is the least linear of the three. At 10 Hz short peaks are missed, and those would push r higher. We still have no measured fuel record of an excavator, so the programme's main test (H4) remains open.

Script and full output: measured_excavator_flow.py, results. The prediction log, with the three misses, is in LOG.md.

All writing · More in Research findings · All tools and research