Platform / Capability
Catch process drift before it costs you.
Fixed alarm limits catch the failures you already know to watch for and miss the slow drift that quietly walks a process out of spec. Manufacturing anomaly detection, run as a vote across four methods, flags the abnormal pattern early and ranks it by what it will cost.
Why anomalies hide until they cost you
Most plant alarms fire on a fixed threshold. A tag crosses a red line, someone gets paged, the line stops. That model catches the failures you already knew to look for. It misses the slow drift that walks a process toward the edge of spec while every single reading still looks fine.
The expensive anomalies are rarely a single spike. They are a bearing warming half a degree per shift, a fill weight creeping toward the low limit, a cycle time stretching a few milliseconds per cavity. None of it trips a threshold. All of it shows up later as scrap, a customer reject, or an unplanned stop.
Fixed limits punish you the other way too. Set them tight and operators drown in false alarms until they mute the system. Set them loose and the real event arrives with no warning. Manufacturing anomaly detection exists to close that gap: to flag the pattern that is abnormal for this process, not just the reading that happened to cross a line.
Four detectors, one vote
The Anomaly Sentry does not lean on any single statistical trick. It runs four independent methods over each signal and treats their agreement as the signal itself. When several methods flag the same window for different reasons, the finding is real. When only one fires, it is usually noise.
- Z-score: measures how far a reading sits from the recent mean, in standard deviations. Fast and clear on stable, roughly normal signals, but sensitive to the outliers that distort the mean it depends on.
- MAD (median absolute deviation): a robust cousin of the Z-score built on medians, so a few wild readings do not move the baseline. Strong on noisy tags where a plain Z-score over-reacts.
- IQR (interquartile range): flags points that sit outside the middle half of the distribution by a set margin. Distribution-free, so it holds up when a signal is skewed rather than bell-shaped.
- Isolation Forest: a machine-learning method that finds the points easiest to separate from the rest across many variables at once. It catches multivariate anomalies no single-tag rule can see, like a temperature and a torque that are each normal but wrong together.
Every method has a blind spot. A Z-score gets fooled by the very outliers it should catch, and a univariate rule cannot see a bad combination of two good readings. Running the four together and requiring a vote is what turns four imperfect detectors into one dependable call, and it is why the same engine works across temperature, pressure, vibration, cycle time, and dimensional data without hand-tuning every tag.
Ranked by dollars, not by alarm count
A detector that surfaces a hundred anomalies a day is not helping anyone. The real question is which two are worth walking to the floor for. The Anomaly Sentry scores every finding on two axes: how confident the ensemble is that the pattern is real, and how much money it is likely to cost if it is.
Confidence comes from the vote. A window flagged by all four methods and sustained across readings ranks above a single-method blip. Dollar impact comes from context: a drift on a bottleneck cell running at capacity is worth more than the same drift on a machine with slack, and an anomaly trending toward a scrap boundary is worth more than one drifting toward a cosmetic limit.
The output is a short, ordered queue instead of an alarm log. Operations sees the next best place to intervene, ranked the way the rest of the platform ranks every opportunity: by dollar impact and confidence, with a human deciding what happens next.
One detector across every signal you already stream
The Anomaly Sentry reads the data your plant already produces. Through 43+ connectors into MES, SCADA, ERP, and historian systems (SAP, Siemens, Rockwell, OSIsoft PI, Ignition, Kepware, OPC-UA, MQTT, Modbus), it watches process and equipment signals side by side rather than in separate silos.
- Process signals: fill weight, dimensional measurements, temperature and pressure setpoints, cycle and dwell times, mix ratios.
- Equipment signals: motor current, vibration, spindle load, hydraulic pressure, bearing temperature.
- Quality signals: in-line gauge readings, first-pass yield, and defect codes tied back to the process window that produced them.
Because the methods are distribution-aware, the same engine adapts to a fast discrete line and a slow continuous process without a rules rewrite. Normalization for shift calendar and product mix keeps a planned changeover or a scheduled slowdown from reading as an anomaly, so the queue stays honest.
Predictive quality, not post-mortem quality
Catching drift early is what turns quality from an inspection problem into a process one. When the Anomaly Sentry flags a fill weight trending toward the low limit three hours before the first reject, the correction is a small setpoint nudge, not a quarantined pallet and a root-cause meeting.
The same early signal separates a quality drift from an equipment problem. A dimensional trend that tracks a rising spindle-load anomaly points at tool wear, not operator error. That link is where the Anomaly Sentry hands off to the specialists beside it: the Quality Sentry for SPC and Cpk context, and the Reliability Forecaster for failure-mode prediction.
This is the practical meaning of predictive quality: shorten the distance between when a process starts going wrong and when someone can act, so the fix lands while it is still cheap and before the loss rolls up into the quality and availability losses that drag OEE down. Modeled against our design-partner program, tightening that loop targets a 5 to 12 percent scrap reduction and 8 to 18 percent less unplanned downtime. Those are modeled target ranges, not results we are claiming for a named plant.
Advisory by design, verified by finance
The Anomaly Sentry never touches a setpoint or stops a line. It surfaces a ranked finding, and a permissioned human decides whether to act. Every recommendation, owner, and outcome is logged and exportable, so the record of what the AI suggested and what the team did stays auditable.
When an action is taken, it does not end at resolved. The change is measured against a normalized baseline and reconciled into the verified savings ledger, so a drift you caught early becomes a number your finance team can trace rather than a story about a save that might have happened.
Data is encrypted in transit with TLS 1.3 and at rest with AES-256, isolated per tenant, and handled under controls aligned with, not certified to, SOC 2 and ISO 27001. You can read the full security posture to see how the detector and its data are governed.
Frequently asked
What is manufacturing anomaly detection? It is software that learns the normal pattern of a process or machine signal and flags readings or trends that depart from it, including slow drift that never crosses a fixed alarm limit. KaizenFlow does this with an ensemble of four statistical and machine-learning methods rather than a single threshold, then ranks each finding by likely dollar impact and confidence.
How is this different from SPC control charts? SPC is powerful and still central to quality, but it is usually applied one characteristic at a time with limits an engineer sets. Anomaly detection runs continuously across many signals at once, catches multivariate patterns no single chart shows, and prioritizes findings by cost. The two are complementary: SPC confirms and documents, anomaly detection scans and prioritizes.
Why use four methods instead of one? Every detector has a blind spot. A Z-score is fooled by the outliers it should catch, univariate rules cannot see a bad combination of two normal readings, and a single model over-fits noisy tags. Requiring several independent methods to agree filters out false alarms and raises confidence that a flagged event is real.
Will it create alarm fatigue? The goal is the opposite. Instead of a raw alarm stream, findings are voted on for confidence and scored for dollar impact, then presented as a short ranked queue. Low-confidence, low-cost blips stay out of the way so the team's attention goes to the few anomalies actually worth acting on.
See it on your signals
Point it at your own tags.
In a walkthrough we run the Anomaly Sentry against a slice of your own historian data and show the ranked findings it surfaces, each scored by confidence and dollar impact.