Most equipment failures announce themselves weeks early through tiny changes in vibration, temperature or current that no fixed alarm limit will catch. Models that learn normal behaviour can flag those subtle departures long before a trip.
Failures are rare, so plants seldom have enough labelled fault data to train a classifier. Learning what normal looks like and flagging departures from it is a more practical approach.
What Is Anomaly Detection?
Anomaly detection is the use of statistical or machine learning models to identify data patterns that differ significantly from learned normal behaviour, signalling possible faults, process upsets or cyber attacks. In plants it extends traditional alarms and supports predictive maintenance.
A study in the journal Sensors on a smart machinery plant reported that its detection rate rose from 85 percent to 96 percent after optimisation, with a false positive rate of 3 percent. A CNN reached 94.2 percent accuracy, ahead of SVM and random forest models.
Data usually comes from the historian, vibration systems and edge devices. The broader AI context is covered in AI and machine learning.
Idaho National Laboratory has studied subtle process anomalies in nuclear plants, where early warning matters most.
5 Powerful Methods Compared
| Method | How It Works | Best For |
|---|---|---|
| Statistical limits and z score | Flags values far from the mean | Single sensors, quick start |
| PCA with T² and SPE | Finds broken correlations between many tags | Multivariable process units |
| Isolation forest | Isolates rare points with random splits | Mixed tabular data |
| Autoencoder | Reconstruction error rises for unusual data | Complex machine signals |
| LSTM forecasting | Prediction error on time series | Dynamic processes |
Most industrial projects use unsupervised or semi supervised learning, because labelled failures are scarce. The model learns only from healthy periods.
For rotating machines, spectrum features feed the model, as in vibration analysis with AI.
How a Detection Pipeline Works
Explanations matter as much as scores. Operators act faster when the alert says which tags drove it.
Models can run on edge devices or central servers, see edge AI vs cloud AI.
Where These Models Help
Network based models complement firewalls in SCADA network security. Vision based inspection is another branch, covered in machine vision inspection.
Controlling False Alarms
Too many false positives create the same fatigue as a bad alarm system. Follow the discipline of ISA 18.2 alarm management and route model alerts to engineers first.
Use persistence rules, such as three consecutive anomalous scores, and retrain after planned process changes. Mark maintenance periods so they are not learned as normal.
Starting a Pilot Project
Start an anomaly detection pilot on one critical asset with good historian data and a clear owner. A compressor, pump set or reactor is usually a better first target than a whole plant.
Measure lead time, the hours or days between the first alert and the confirmed fault. That number is what convinces management to scale the solution.
Z Score Formula
Example, bearing temperature:
Normal mean 62 °C, standard deviation 1.5 °C
Current value 67.1 °C
z = 5.1 ÷ 1.5 = 3.4
Above a threshold of 3, so flag as anomalous
A z score threshold of 3 flags about 0.3 percent of normal points if the data is roughly normal. Multivariable methods catch faults that single tags miss.
Z Score Calculator
Recalculate mean and deviation per operating mode, because startup and full load behave differently.
- Early fault warning.
- No need for failure labels.
- Covers many tags at once.
- Finds unknown problems.
- False positives and alert fatigue.
- Changing operating modes.
- Data quality issues.
- Need for clear explanations.
INL Process Anomaly Detection PDF
Anomaly Detection With MATLAB Video
Anomaly Detection FAQ
Related Articles
- Vibration Analysis With AI
- Predictive vs Preventive Maintenance
- AI and Machine Learning
- AI in PLC, SCADA and DCS
- Digital Twin in Automation
External References
- Subtle Process Anomalies Report, Idaho National Laboratory
- Smart Industrial Machinery Study, Sensors Journal
- Anomaly Detection, Wikipedia
What We Learn Today
- Anomaly detection learns normal behaviour and flags departures.
- PCA, isolation forests and autoencoders suit different data.
- Persistence rules and clear explanations limit false alarms.