Anomaly detection

Anomaly detection: how machine learning finds rare patterns in consumption, process and payment data, which methods make it work and in which fields the investment pays off.

The most expensive mistake
is the one nobody sees

It drowns in the normal case: one wrong number among a thousand plausible ones. Anomaly detection turns that around and lifts the rare pattern to the surface before it gets expensive.

What it does

What anomaly detection delivers, and what it does not

Anomaly detection reliably answers a single question: which value does not fit what this process normally does? It does not answer why. An alert is a finding, and the next step is always human: is the outlier a sensor fault, a mixed-up unit or a genuine rare event? Systems that skip this classification produce alert fatigue instead of insight.

That is why we never build anomaly detection as a black box, but with traceable thresholds and a verification path per alert. We have published the methodology openly: our open-source data quality audit uses the same techniques to find suspicious values in datasets before a model is built.

The seven-question audit: the same methodology, applied to data quality

Fields of use

Where anomaly detection pays off

The technique is always the same: learn the normal behaviour of a process and flag deviations. The value appears where deviations cost money and nobody is currently looking.

Fully calculated: consumption anomalies in a fleet's diesel usage

Model drift and data drift: monitoring in operation (phase 6)

Consumption and fleets

Fuel, energy or material consumption per vehicle, machine or order: anomalies reveal defects, misuse and shrinkage long before the monthly report smooths them away. We have published a fully calculated diesel consumption case.

Quality in manufacturing

Process values drifting out of their settled state announce scrap before inspection counts it. Anomaly detection on process data is the earliest quality indicator available.

Energy and peak loads

Unexpected consumption patterns in load profiles reveal faulty equipment, forgotten consumers and expensive simultaneity that stays invisible in aggregate meters.

Fraud patterns in bookings and payments

Fraud detection is, methodically, anomaly detection: rare patterns in the normal noise of many transactions. The same techniques flag suspicious bookings, duplicates and unusual payment routes for human review.

Models and data in operation

AI systems themselves need monitoring: when input data drifts, models degrade silently. Drift detection is anomaly detection on the meta level and belongs to every model operation.

Methods

Anomaly detection with machine learning: the methods behind it

The foundations are older than the AI label and reliable precisely because of it. Statistical outlier rules such as Tukey's draw the line at 1.5 times the interquartile range, stricter thresholds at five standard deviations. Distribution comparisons such as the Kolmogorov-Smirnov test recognise when a process shifts as a whole. Machine learning enters where normal behaviour is too complex for fixed thresholds: models learn the relationship between context and expected value, and the anomaly is the deviation from that expectation.

The method follows the question, not the other way round. For many cases robust statistics with explainable thresholds is enough, and that is exactly where a serious project starts, before anyone proposes a neural network.

Which of your processes should go first is settled by our AI strategy consulting from first conversation to roadmap . If the answer is robust statistics instead of AI, we say so.

Frequently asked questions about anomaly detection

What is anomaly detection, in short?

A technique that learns the normal behaviour of a process from data and flags values that do not fit: the consumption spike, the unusual booking, the process value outside its settled state.

Which methods does anomaly detection with machine learning use?

The basis is statistical outlier rules (Tukey criterion at 1.5 times the interquartile range, five-sigma thresholds) and distribution comparisons such as the Kolmogorov-Smirnov test. Machine learning complements them where normal behaviour depends on context: the model learns the expected value, and the anomaly is the deviation from it.

Is fraud detection the same as anomaly detection?

Fraud detection is a field of application of anomaly detection: rare, atypical patterns in booking and payment data. The methodical core is identical. The difference lies in the process afterwards, because a fraud suspicion needs a regulated human review, not an automatic verdict.

How much data does anomaly detection need?

Enough to describe normal behaviour, which is often less than expected: for statistical thresholds a few months of clean process data are frequently sufficient. Data quality matters more than volume. Whether your current data carries a project is testable before it starts.

What happens when the system raises an alert?

The alert is a finding with a verification path: which value, which threshold, which context. The classification stays human, because whether it is a sensor fault, a mixed-up unit or a genuine event is decided by domain knowledge. Good anomaly detection reduces review work to a few justified cases instead of a hundred unfiltered alarms.

Service

Let us talk about your data

Free 30-minute initial call. Briefly describe your dataset and your plan; optionally you can state a preferred date.

Free 30-minute initial call
Fixed price at a clear scope
Your data stays with you
Result as a management report

These services do not replace legal data protection advice.