A histogram is a bar chart of how often measured values fall into ranges, so it shows the shape, center, and spread of a process at a glance. In quality control you read three things from it: the shape of the distribution, where it is centered, and how wide it spreads against the specification limits.
A column of numbers hides its own story. Sorted into bars, the same data reveals whether the process is centered on target, whether it is tight or loose, and whether something odd, two peaks, a long tail, a wall on one side, is going on underneath. The histogram is one of the seven basic quality tools for exactly this reason: it turns a table nobody can read into a picture anybody can. This guide covers how to build one, how to read its shape, what the common shapes mean, and how it connects to process capability and statistical process control.
What is a histogram in quality control?
A histogram groups measured values into equal-width intervals, called bins, and draws a bar for each bin whose height is how many values landed there. Unlike a bar chart of categories, the bins are ordered along a continuous scale, say, shaft diameters from 9.90 to 10.10 millimeters, so the bars sit side by side and their outline traces the shape of the data. That shape is the whole point: it is the distribution of the process made visible.
In quality work the histogram answers a question a single average cannot: not just where the process sits, but how it behaves. Two processes can share the same mean and be wildly different, one tight and centered, one spread out and lopsided, and only the picture shows it. Mark the specification limits on that picture and you can see immediately what fraction of parts fall outside the requirement, which is why the histogram is a first stop before any capability calculation.
How do you build a histogram?
You need a set of measurements, usually at least 50 to 100 for the shape to settle, and a few decisions about how to bin them. The steps are straightforward:
- Gather enough data. Collect at least 50 to 100 measurements of one characteristic from a stable period of production. Too few values and the shape is just noise; a handful of points cannot show a distribution.
- Find the range. Subtract the smallest value from the largest to get the spread the bars must cover. This sets the width of the whole chart.
- Choose the number of bins. Pick roughly the square root of the number of data points, commonly between 5 and 20 bins. Too few bins hide the shape; too many break it into a spiky mess.
- Set the bin width. Divide the range by the number of bins and round to a convenient number, so every bar covers an equal interval. Equal width is essential, unequal bins distort the picture.
- Tally the values into bins. Count how many measurements fall in each interval. A tally sheet or a spreadsheet frequency function does this quickly.
- Draw the bars and add the limits. Plot each bin's count as a bar, then mark the lower and upper specification limits and the target on the same axis so shape and spec appear together.
- Read shape, center, and spread. Step back and read the three things that matter: what shape the outline makes, where the peak sits relative to target, and how the width compares to the distance between the limits.
By the numbers. The histogram is one of the "seven basic quality tools" recognized by the American Society for Quality, used to display the distribution of data so patterns that hide in a table become visible (ASQ, Seven Basic Quality Tools). ASQ describes the histogram specifically as the graph for showing frequency distributions and judging whether a process is centered and capable against its limits (ASQ, What is a Histogram?). For binning and shape interpretation, the NIST/SEMATECH engineering statistics handbook's exploratory data analysis section documents how bin count affects the picture and how to read skewness and multiple modes (NIST/SEMATECH e-Handbook, Histogram). Guidance on sample size and bin count varies, treat the 50-to-100-point and square-root-of-n rules as common practice, not fixed law.
How do you read the shape of a histogram?
Read the outline first, then the position, then the width. The outline is the shape of the distribution, a single centered hump, a lopsided lean, two peaks, a flat top. Each shape is a clue about the process behind it. A smooth, symmetric bell is what a stable, single-cause process usually produces, so anything that is not a bell is worth a second look. The shape is asking you: is one consistent process at work here, or is something mixed in?
Then read position and width against the specs. Position is where the peak sits relative to the target and the limits, dead center, or pushed toward one side. Width is how far the bars spread compared to the gap between the lower and upper limits. A narrow distribution sitting on target is a healthy process with room to spare; a wide one, or one shoved against a limit, is trouble even if the average looks fine. Those three reads, shape, center, spread, are the whole discipline of histogram analysis.
What do common histogram shapes mean?
Certain shapes recur and each points at a cause. A bell (normal) shape means a stable process driven by many small random causes, the baseline healthy pattern. A skewed shape, with a long tail to one side, often reflects a natural boundary: a dimension that cannot go below zero, or a process bumping against a physical limit, will pile up on one side and tail off the other. Skew is not always a problem, but it tells you the process is not symmetric and a plain average and standard deviation may mislead.
The shapes that most often signal trouble are the mixed ones. A bimodal shape, two distinct peaks, almost always means two populations are blended in the data: two machines, two shifts, two lots, two operators, or two gauges, each centered differently. The fix is to stratify: split the data by suspected source and re-plot, and the two peaks usually resolve into two clean distributions you can address separately. A truncated or cut-off shape, where the distribution ends abruptly at a wall, often means the parts were already sorted or the gauge cannot read past a point. A comb shape, with alternating tall and short bars, usually points at rounding or a resolution problem in the measurement, not the process.
How do you read a histogram against specification limits?
Draw the specification limits and the target on the histogram, then compare the distribution to them. There are two failure modes and they call for different fixes. If the distribution is narrow enough but shifted off to one side so it crosses a limit, the process is off-center, you adjust the setpoint to move it back onto target, and the tail comes back inside. If the distribution is centered but so wide that both tails cross the limits, the process is too variable, no amount of centering fixes it, and you have to reduce the spread by attacking the sources of variation.
This centering-versus-spread read is exactly what a capability index puts a number on, but the histogram shows it to you first and more honestly. A capability number can look acceptable while the picture reveals a skew or a second peak that the index quietly assumed away. Always look at the shape before trusting the index, a Cpk computed on bimodal data is meaningless, because the math assumes a single distribution that is not actually there.
How does a histogram relate to capability and SPC?
The histogram is the front door to process capability. A capability study asks how the process spread compares to the spec width and how well centered it is, the same two questions you read off the histogram by eye. The picture is the sanity check on the number: if the histogram is a clean, centered bell inside the limits, a good Cpk is believable; if it is skewed, bimodal, or truncated, the capability index is built on a false assumption and needs a harder look. Never report a capability number without looking at the shape it came from.
The histogram and the control chart are partners, not substitutes. A histogram is a snapshot, it shows the distribution but throws away the time order, so it cannot tell a stable process from one that drifted across the sampling window. A control chart keeps the time order and shows stability but not the full shape. Read them together: the chart says whether the process is in control, the histogram says whether an in-control process actually meets the specs. And both depend on trustworthy data, a histogram of readings from a biased gauge is a picture of the gauge as much as the process, which is why a gage bias and linearity check comes first, alongside the rest of the seven basic quality tools like the Pareto chart.
The practical catch is that a histogram is only as timely as the data behind it, and a chart built at month-end from a stack of clipboards tells you about a process that already ran. When measurements are captured live at the point of inspection, the distribution builds in real time and a shift toward a limit or a second peak appearing shows up while the run is still going, early enough to adjust instead of explain. That live feedback is what Harmony gives a plant, turning a picture drawn after the fact into one the floor can watch as it forms. CLS made that shift, from measurements found the next morning to measurements visible during the shift.