How CureWood replaced a 1991 Siemens S5 in one 8-hour cutover

01This is not a simple machine

CureWood Engineered Building Products makes finger-jointed lumber in Surrey, British Columbia. Ben McKinnon started the company to do something specific with cheap wood: “take smaller, low-value pieces of wood and turn them into more advanced, higher-value pieces of wood.”

The machine that does it is a GRECON FJ finger-jointing line commissioned in 1991, and it is a serious piece of equipment. A magazine conveyor stages packages of boards. Two cutter stations mill interlocking joint profiles, one for each end of the board, each with its own clamps, vibrating tables and feed drives, handing work across a transfer table between them. Glue goes on, feed rollers drive the boards end to end through a length-measuring system, a saw cuts the moving string at exactly 2.50 meters, and crossfeeds push the cut pieces into a hydraulic press. One controller choreographs all of it through 98 sensors, 91 outputs and 54 timers. When the line is running well, each cutter table cycles every 23 seconds and the press ejects a finished charge every couple of minutes.

That controller is a Siemens SIMATIC S5, programmed in May 1991 and last modified by hand in November 1994. Ben had to recover the original program from the machine builder in Germany. The rest of the documentation is paper: the ladder printout that shipped with the machine, part German and part English, and a wiring schematic he described as a full phone book, which he took to a Staples self-service scanner and emailed over in two parts. This is normal for how CureWood grew.

Every change to the line ran through that paper and a laptop plugged into the rack. The risk was not the work. It was what the work might cost him.

02Nobody quotes a 30-year-old machine sight unseen

Before we priced our own work, we tried to get a market baseline. Our research suggests similar projects typically run $100k–$150k. We worked with Ben on pulling competing quotes, and what he said is the reason this genre of project stalls for years:

We believe in price transparency. We built the quote bottom-up from fundamentals, and we charged him:

• $2,950 for the Cortex AIC
• $3,000 in third-party I/O hardware
• $4,000 for integration hours and to cover our international travel

03Step one was reading a program nothing modern can open

The S5's program lived in two places: on that paper, and as MC5 machine code inside the rack. We scanned every page and put twelve agents on transcribing them in parallel, with one hard rule: flag anything smudged or ambiguous instead of guessing. Then we decoded the machine code itself, 2,420 instruction words across 22 blocks, with a decoder we could prove correct two ways. Re-assembling our decode reproduced all 2,420 statements exactly, and an independent open-source Step5 tool read 31 of 32 blocks the same way we did.

Good thing we did both, because the paper is wrong in places. Simulation turned up five printed contact readings the machine physically could not run under, and all five were spots the transcribers had flagged.

Then Ben sent us the program actually running in his rack, with a note that it was the 1991 program “with our small little edits.” Those edits mattered. The running program differed from every document we had in 15 places, including six timer presets somebody had retuned over thirty years and never written down. We built to the machine, not to the paperwork.

04We proved the port before we touched the machine

We translated the program, network by network, into real-time Python on our controller, the Cortex AIC. We also built a behavior simulation of the whole line into an FMU model: the magazine, both cutters, the transfer, the saw, the press. Then we ran the ported program against it until the simulated line made lumber the same way the real one does.

Then came the real test. We checked our port against the program running in the rack, statement by statement, all 300 of them. Zero differences. That is the whole basis for what we told Ben in writing, three days before we flew: the schematics had anchored the simulation model, and we felt really confident about one-shot success of the retrofit. Every retrofit promises the new system will behave exactly like the old one. We were willing to put it in an email first.

05What the calendar actually looked like

06The cutover took one shift

We arrived on site, Cortex in hand, with the new EtherCAT I/O terminal. 189 wires, one 8-hour shift from lockout to power-up, and most of that was screwdriver time.

Two things we deliberately did not touch. The hardwired safety chain of e-stops, guard circuits and motor starters was never part of the PLC program and stayed exactly as it was. And the S5 rack came out intact. Every wire lands back where it came from, so the old controller is still a physical rollback path. CureWood runs on our stack now, and supporting it is our job.

07Then we turned it on

The hydraulics came up. Ben went through the machine the way an owner does. He cycled every operator console. He jogged every moving component by hand. Then he ran real boards through end to end: out of the magazine, profiles milled at both stations, glue on, the saw cut, a pressed charge out the back. He checked the finished product and went feature by feature, looking for anything that behaved differently. After thirty minutes he ran out of things to check.

One shot, exactly as we called in the email.

08What the line gained

The logic is unchanged on purpose. Everything around it moved 35 years forward. Right after power-up we built new HMI panes that Ben can draw and maintain by himself. The line now has a historian recording every one of the 189 points, remote access, and a program a human can actually read. Neuron's AI agent is now also live with CureWood, able to answer troubleshooting questions based on data and build new logic in plain English.

The historian went to work immediately on the faults that used to be unfindable.

It proved itself during commissioning on exactly that kind of fault: a photoelectric sensor that confirms one of the moving carriages is in position started reading marginal. The trace showed it right away, and the diagnosis took minutes. Dust on the lens. One wipe and the signal came back. On the S5, that would have been an afternoon of wire wiggling.

Watch · live troubleshooting

09This is only the beginning

Three days after the line was running, we sent Ben a proposal for two instrumentation clusters, each in a small cabinet near the sensors it serves: one by cutter 2 and the glue dispense, one at the end of the line. Coupler, analog input card, digital in and out, a string potentiometer for position, and a camera for the glue work. He approved it the same morning.

Those two clusters are the upgrades he had wanted for years and could not have on the S5. The first adds carriage position feedback and patterned glue dispensing on the finger joints, so the press cannot over-squeeze a joint. The second retires the most hated part on the machine, the length encoder: today a wheel rides along the wood, glue keeps fouling it, and it dies regularly. A string pot on the staging-table backstop measures the same thing without ever touching the product. He is already sending us photos to train the glue-inspection model, and offering to build the camera mount and lighting himself.

If you have a machine running on a controller nobody supports anymore, hit us up.