Last October, at 9:47 on a Tuesday morning, our Keyence color sensor started throwing rejects on the injection molding line. One out of every four parts got flagged. The operators picked them up, squinted, and saw nothing wrong. I did the same. My first thought — my wrong thought — was that the sensor had gone haywire.
I've been handling process and QC for a small contract manufacturer for about seven years now. I've personally made (and documented) enough mistakes to fill a binder, and this one nearly made it to the front cover. We don't have a dedicated QC department. I'm the process engineer who somehow became the person responsible for deciding whether a rejected part actually goes in the scrap bin. Our runs are small — a few thousand pieces at most — but they go into medical and automotive assemblies, which means the cost of a wrong decision is high.
That morning we were running 1,200 molded components for a medical customer. The rejection criterion was "off-color or blemished surface," a leftover from a past complaint. That's why we'd installed the Keyence color sensor in the first place: to catch what human eyes miss.
Tuesday Morning: The Color Sensor Rejects Everything
The operator's first instinct was to dial down the sensitivity. It's not a dumb instinct. If a sensor rejects parts that look fine, and the customer's drawing doesn't mention a "faint film," the sensor seems overzealous. In my first year at this plant, I did exactly that: turned down the threshold, let the parts run, and shipped a batch of bad product. Learned that lesson the hard way. So this time I asked him to wait while I ran a few checks.
I grabbed the No. 436 micrometer out of the tool drawer — the one that's been there since before I started — and checked the critical outer diameter on the last ten rejected parts. All within spec. Dimensions weren't the trigger.
Then I remembered the Keyence digital microscope sitting in the corner. I'd been reading the Keyence microscope news when they launched the new VHX series, and I convinced my boss to buy one for exactly this kind of failure analysis. Ten seconds under the microscope and I saw it: a faint film on the surface of the rejected parts. The naked eye couldn't see it. The color sensor could.
The sensor wasn't lying. It was catching something we couldn't. (Note to self: stop assuming the sensor is the first thing that's wrong. It's usually not.)
The HPLC Pump Pressure Was the Clue
Now I needed to know what the film was. It wasn't mold release. It wasn't grease. I swabbed a few rejected parts and ran them through our Agilent HPLC system, which we use to verify that there are no process residues on parts we sell to medical customers.
This is where the story gets embarrassing.
Our HPLC pump had been creeping up in pressure for weeks. Retention times on the calibration standard had drifted. I had literally Googled "when to change your columns hplc agilent" and read the standard advice: rising backpressure, drifting retention times, peak tailing, loss of resolution. We had three of those signs. And I chose to ignore them.
Why? Because we're small. A replacement analytical column is a real line item when your biggest job is 1,200 pieces. I told myself the column had a little life left. Looking back, I should have ordered the replacement the first time the retention time drifted. At the time, it felt like a reasonable way to stretch the budget. It wasn't.
The data exposed me quickly. The first swab sample showed a contaminant peak. The re-run of the same sample gave a completely different result. Same sample, same method, different answer. That inconsistency was the smoking gun — and I still spent half a day suspecting the pump, the detector lamp, and the injection valve. Everything except the column.
Finally I called an applications engineer at Agilent. Her first question: "How old is your column?" Nine months and roughly a thousand injections, many of them unfiltered swab samples. We ran a system suitability test and failed it in two categories. I stopped the line, swapped in a new column (and added a guard column this time), and re-ran the swabs.
The parts were clean. The Keyence color sensor had been right the entire time.
I nearly let a $600 column destroy a $25,000 order and a customer's trust. The sensor cost us nothing. The column cost us plenty — because I was too cheap to replace it when the data said to.
When to Change Your HPLC Columns (Agilent or Any Brand)
I don't want you to repeat this mistake, so here's the boring, practical version now printed on a card above our HPLC:
- Watch the backpressure. If the HPLC pump pressure climbs more than 10–15% above its normal baseline at the same flow rate and mobile phase, the bed is packing down or the frit is clogging. Both mean the column is living on borrowed time.
- Watch the retention times. If your standard's retention drifts by more than about a tenth of a minute run-to-run, or shifts consistently day over day, the column chemistry has changed. This was our first warning sign.
- Watch the peak shape. Tailing, fronting, or split peaks after months of good performance mean the bed is damaged or contaminated.
- Watch the replicates. If the same sample gives different results on the same method, check the column before you blame the pump. In my case, the pump was innocent.
- Use a guard column. It's far cheaper to replace a guard cartridge than a full analytical column. We now replace the guard every 200 injections, or sooner if the samples look dirty.
- Keep a column log. Date, number of injections, mobile phase, backpressure at start and end. When you're unsure whether the column is done, the log decides.
Agilent's own column care documentation on agilent.com lists these same symptoms, and other manufacturers publish similar guidance. The answer to "when to change your columns" is not "after X months." It's "when the data says so" — and you'll only know that if you're tracking the data.
What This Small Shop Learned
A few weeks later, I told this story to a friend who runs a small machine shop. He laughed and said, "You actually called Agilent? I figured companies like that only care about big accounts." I used to believe that too.
The idea that big vendors ignore small customers comes from an era when field reps were routed by annual sales volume. Today, that's changed. The Agilent applications engineer spent an hour with me on a column we'd bought through a distributor. And the Keyence support rep who helped me reconfigure the color sensor's tolerance window didn't ask how many parts we run per year. He asked what we were trying to detect, then walked me through the sensor settings for about 40 minutes. Nobody told me we were too small for that.
The myth that big vendors ignore small customers is a dead one. If you are a small shop, a startup lab, or a plant running prototype quantities: call the vendor. Ask the question. The call costs nothing, and the answer might save you a whole lot more.
Small doesn't mean unimportant. It means you can't afford to waste money on rework, rushed shipping, and repeat testing — which is exactly what I did by trying to squeeze one more month out of an HPLC column that was already gone. Trust the data. Replace the consumable. Call the support line. That's the whole lesson.
