Can you find a mill's bottleneck from the data alone?

Usually, yes, if the machines log their states with timestamps: running, down, blocked by the machine after, starved by the machine before. Count a station as active when it is running or down, including repair and changeover, and inactive when it is blocked or starved. The bottleneck is then the station whose active periods run longest with the fewest interruptions. It is not always the station with the most downtime. A stop anywhere spreads along the line as blocked and starved time, so a downtime report alone often points at a machine that is suffering the bottleneck, not causing it. The bottleneck can also move between shifts, products and log mixes.

What the usual answer says

The usual advice is to look for the machine with the lowest throughput, the most downtime or the longest queue in front of it, and to track it on an OEE dashboard. Those are reasonable first looks. Each can mislead in a line where machines are coupled by small buffers.

What the research says

A systematic review by Anders Skoogh, Christoph Roser and colleagues, published in November 2023, screened 412 articles and analysed 27 in depth. The authors' open copy is free to read. It identified 14 bottleneck detection methods, classified by whether they use queue states, process states, or both. It also identified three ways of putting them into practice: walking the floor, simulation and data science. "Most recent literature focuses on the development of data science approaches."

The review explains why simple measures mislead: "every time a random event occurs on a machine, its effects may be propagated to the upstream and downstream machines in the form of blockage and starvation". The machines around a constraint spend time idle because of it. On walking the floor, it makes two points. Walking the line is time-consuming and "may be challenging with machines that have extremely small cycle times and in environments where the throughput bottleneck shifts across machines frequently". And "human observations are prone to errors and can lead to wrong identification of throughput bottlenecks".

It describes one well-tested method in plain terms. What counts as active matters: the review notes that the original definition "should include all activities towards increasing the production system throughput, such as repair, service and set up activities". A station that is down for repair is active in this sense, because it is the one holding the line up. Under the average active period method, "the station with the longest average active period is considered to be the average bottleneck, as this station is the least likely to be interrupted by other processes and thus dictates the overall system output".

What it looks like in a mill

From Quarri's own work with a sawmill: the trim line was the bottleneck, down 75%+ of the time. Here the downtime report and the constraint agreed. In other mills they don't: an edger that is often blocked by a slow trimmer shows plenty of idle time, and speeding it up changes nothing. The state logs tell the two apart. Blocked time on the edger, and starved time on the sorter, both point at the trimmer between them.

Shifting matters too. Different log mixes and cutting patterns load stations differently. As an example, the headrig might constrain the mill on large logs and the trimmer on small ones. The review notes that shifting bottlenecks come partly from "the actions taken by practitioners to resolve the throughput bottleneck".

A caution for sawmills

The review's examples are mostly serial production lines, where one part moves from machine to machine. A sawmill splits: one log becomes a cant, then many boards, and transfer decks between stations hold varying amounts. We know of no study that tests the active-period method on a sawmill line. Our view is that it transfers if states are read at each station as they are, and if the comparison between stations is weighted by volume or pieces handled, since a station that runs long on few pieces is different from one that runs long on many. Test it against a period where the constraint is already known, as in the example above, before relying on it.

What data you need

The minimum is a timestamped state for each main station: running, down (with a reason if possible), blocked and starved. Some sawmill control systems already record machine states for their own purposes; check what yours holds. The work is getting them out, aligning the clocks and labelling blocked and starved correctly, since some systems record only on and off.

With a week or two of states, compute each station's share of active time (running plus down) and its average uninterrupted active period, by shift and by log class. The station that leads on both is the average bottleneck. Then look at how often, and when, another station takes over.

What AI adds

AI helps with the preparation more than the method: reading reason codes typed by operators into consistent categories, aligning logs from different systems, and flagging gaps. It can also watch the states live and say when the constraint has moved, which a monthly report cannot.

When it doesn't apply

Mills whose machines log only running and stopped, with no blocked or starved state, need to add that first, or infer it from buffer sensors. Very simple lines with one dominant machine may not need the analysis. And a constraint set by the market, such as orders for only one product, will not show up in machine data at all.

Quarri for sawmills is built around how a sawmill runs, from log intake to shipped order.

Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.