A line is specified at 60 parts an hour, passes its acceptance test at 59, and settles at 48 once it is in production. Nothing is broken. The difference is a chain of small losses that every line has and almost no specification names, and each one of them can be calculated before the machine is built.
The complaint arrives in roughly the same words every time. The line was bought for 60 parts an hour. It was tested at the builder, it made 59, everybody signed. Six months later the shift report says 48, the maintenance log is almost empty, and no single station can be shown to be at fault. Somebody suggests the machine has worn, somebody else suspects the operators, and a meeting is booked to discuss a problem that nobody has yet measured.
In almost every case we have been asked to look at, the machine is doing exactly what it was built to do. What is missing is the arithmetic between the cycle time of the slowest station and the number of good parts that leave the line in a shift. That arithmetic has five or six terms, every one of them ordinary, and together they take a quarter off the output. It is worth doing on paper before the steel is cut, because afterwards each term costs money to recover.
Takt, cycle and the number in the contract
Three numbers get called cycle time and they are not the same. Takt time is what the customer demand asks for: available production time divided by the parts needed. If a plant needs 480 parts in a shift that has 480 minutes of attendance, takt is 60 seconds. Station cycle time is how long one station takes to do its work. Line cycle time, on a synchronous machine where everything moves together, is the longest station cycle plus the time the transfer itself takes.
Take a six-station rotary assembly machine with station times of 46, 52, 58, 44, 55 and 49 seconds, and an index that takes 3 seconds to move the dial and settle it. The line cycle is 58 + 3 = 61 seconds, which is 59 parts an hour, not 60, and that is before anything goes wrong. The five stations that finish early contribute nothing: they wait. A machine of this type runs at the speed of its slowest station and pays the index on top, which is why balancing stations to within a few seconds of each other matters more than making any single one fast.
That also shows where to spend money. Taking four seconds out of the 58-second station lifts the line to 62 parts an hour. Taking twenty seconds out of the 44-second station changes nothing at all. We have watched a great deal of engineering effort go into stations that were never the constraint, usually because they were the interesting ones.
Every station you add multiplies the losses
Stations in series share their failures. If each of six stations is independently available 99 per cent of the time, the line is available 0.99 to the sixth power, or 94.1 per cent. At 98 per cent per station it is 88.6 per cent. Nothing is wrong with any station; the arithmetic simply stacks. On 61 seconds of line cycle, 88.6 per cent availability is 52 parts an hour.
This is the strongest argument there is for not adding stations casually. Every extra operation, every extra sensor that can be blocked by a chip, every extra pneumatic cylinder with a reed switch that can drift, enters the product. A function that could be done by a geometric feature in the fixture instead of by an actuator is worth real money over the life of the machine, and it never appears in a comparison of two quotations.
Micro-stops are not stops
A maintenance system records breakdowns. It almost never records the stop that lasted ninety seconds because a feeder bowl jammed and the operator cleared it without telling anyone. Those are the losses that explain the gap between 59 and 48, and the way to find them is to count them.
Suppose one vibratory feeder jams once every 150 parts and takes 90 seconds to clear, including the time before anybody notices. That is an average of 0.6 seconds added to every cycle, which does not sound like much until the second feeder does the same, the screwdriver cross-threads once every 400 parts, and a part arrives at the camera station in the wrong orientation once every 250. Four such events give about 2 seconds a cycle, which on a 61-second line is another 3 per cent. The arithmetic is simply stop duration divided by parts between stops, summed over everything that stops.
Feeders deserve a paragraph of their own. In our experience most micro-stops on an assembly machine come from the parts, not from the machine: a burr left by a worn punch, a stamping that nests into another stamping, a moulded part whose gate vestige is a tenth of a millimetre longer this week. The machine was proven with a box of good samples and then meets a production lot. This is why we ask for parts from the ordinary production run, including the ones at the limits of the drawing tolerance, before a feeder is designed, and why a specification that says nothing about incoming part quality is an argument waiting to happen.
Quality losses come off the top
Only good parts count. A line running at 52 an hour with 1.5 per cent scrap and 2 per cent of parts going to a rework bench delivers about 50 good parts an hour, and the rework bench is a cost of its own that nobody put in the business case. Rework also hides a trap: the part that comes back through the line a second time occupies a cycle, so a 2 per cent rework rate that is reintroduced upstream costs 2 per cent of the line capacity as well as the labour.
The shift is shorter than the shift
The last term is the calendar. An eight-hour shift has 480 minutes of attendance. Take off two breaks of 15 minutes, 10 minutes of startup checks and first-part approval, 10 minutes of shutdown and cleaning, and 15 minutes for one changeover, and 420 minutes remain. That is 87.5 per cent of the shift, and on 50 good parts an hour it gives 350 parts, not the 480 the takt calculation promised.
Multiplied out, the chain reads: 59 parts an hour nominal, 88.6 per cent availability, 97 per cent performance after micro-stops, 96.5 per cent quality, 87.5 per cent of the shift actually available. The product is about 48 parts an hour of good output against attendance. Nothing in that sequence is a fault, and every figure in it is a decision somebody made, usually without being asked.
Where a buffer buys the losses back, and where it does not
Buffers decouple failures, which is why a line of linked but independent stations with small buffers between them usually beats a rigid rotary machine of the same nominal speed. The arithmetic of a buffer has two parts, and people usually do only the first. The size has to cover the stop: a 90-second stop at a 30-second cycle needs three parts of buffer to keep the downstream fed. But refilling that buffer afterwards needs spare speed upstream, and if every station runs at the same cycle time there is none: the buffer empties once and never recovers.
So a buffer is only worth building when the stations around it have a few per cent of speed margin, which means designing the non-constraint stations slightly faster than the constraint rather than balancing everything to the same number. On lines where this was done we see buffers of three to five parts recover most of the availability loss. On perfectly balanced lines the same buffers do nothing at all, and we have been asked more than once to explain why.
Measure, then argue
All of this is arithmetic on assumptions until somebody timestamps parts. The cheapest instrumentation on any machine we build is a record, per part and per station, of when the part arrived, when the station finished, and why the machine waited. That record costs almost nothing in a modern controller and it ends every argument about where the time goes.
What matters in the data is the distribution, not the average. A station whose mean cycle is 52 seconds and whose 95th percentile is 71 seconds is the constraint one cycle in twenty, and an average will never show it. The usual causes of a long tail are a sensor that needs a second look, a feeder that occasionally presents a part badly, or a torque step that retries. Sorting stop reasons by frequency multiplied by duration, rather than by frequency alone, puts the week of engineering effort where it returns the most.
What we write into a specification
For a line we build, output is stated as good parts per hour over a continuous run of a defined length, together with the assumptions it rests on: the quality of incoming parts, the air pressure and the electrical supply at the machine, the presence of an operator, and which stoppages are excluded. A number without those conditions is not a guarantee, it is a sentence.
At the acceptance test we run two figures, because they answer different questions. A short run of an hour or two shows the machine can reach its nominal cycle. A continuous run of a full shift with the customer parts, counting everything that happens, shows what the plant will actually see, and it is the only number worth planning capacity on. The gap between those two figures is the subject of this article, and when it is measured at the builder rather than discovered six months later, it is still cheap to close.
— GANI Engineering engineering team