Anomaly detection is the identification of observations or patterns that deviate significantly from normal behavior in data. In industry it flags unusual sensor readings, process states, or product images that may signal a fault or defect.
Most industrial anomaly detection is unsupervised or semi-supervised. The model learns what "normal" looks like from historical data, then scores new data by how far it departs. This suits plants because labeled failures are rare. Methods range from statistical limits and control charts to clustering, isolation forests, autoencoders, and forecasting models that flag large prediction errors.
Anomalies come in three types. Point: one reading far off (a temperature spike). Contextual: normal in one situation, abnormal in another (high motor current during idle). Collective: a sequence that is unusual as a whole even if each point looks normal.
An anomaly is not a failure. It is a signal that something changed. It may be a sensor fault, a new product, a recipe change, or a real problem. Without context from MES or maintenance records, alerts pile up and operators stop trusting them. As the NIST/SEMATECH handbook notes for outliers, they should be investigated, not simply removed.
Questions about Anomaly detection
Does an anomaly mean a failure?
No. An anomaly is a signal that something changed. It may be a sensor fault, a new product, a recipe change or a real problem. As the NIST/SEMATECH handbook notes for outliers, they should be investigated, not simply removed.
What are the types of anomalies?
There are three. A point anomaly is a single reading far off, such as a temperature spike. A contextual anomaly is normal in one situation and abnormal in another, such as high motor current at idle. A collective anomaly is a sequence that is unusual as a whole.
Does anomaly detection need labeled failure data?
Most industrial anomaly detection is unsupervised or semi-supervised. The model learns what normal looks like from historical data, then scores new data by how far it departs. This suits plants because labeled failures are rare.
Anomaly detection in real plants
- Toast and Instrumental: Toast installed 6 Instrumental stations across 2 hardware programs to inspect each assembly stage of its restaurant point-of-sale hardware at its factory partners.
- Ørsted and ABB: Ørsted is piloting ABB's AI-based fault prediction at Avedøre Power Station: data from ABB protection relays and the centralized control system is analyzed to spot early signs of electrical faults.
- Aker BP and Cognite: Aker BP launched a program to deploy AI-driven anomaly detection on equipment such as wells and heat exchangers, feeding data from Cognite's platform into NVIDIA NV-Tesseract models.
- Gebr. Baldauf and MVTec Software: The dairy Gebr. Baldauf uses a mobile cheese-care robot from Eberle whose 4K camera and MVTec HALCON deep learning inspect ripening cheese wheels for mold and defects.
- Stellantis and Tractian: Stellantis put Tractian energy monitoring sensors on welding robot control panels at one plant to track each robot's electrical behavior; the project is being scaled to 134 machines.
Vendors linked to Anomaly detection
AVEVA (3) · Instrumental (3) · Cognite (2) · ABB (1) · Basler (1) · Boston Dynamics (1) · Hitachi (1) · MVTec Software (1) · Siemens (1) · Tractian (1) · Uptake
Related terms
Condition monitoring
Condition monitoring is the ongoing measurement of machine parameters such as vibration, temperature, or oil quality to detect significant changes that indicate a developing fault.
Predictive maintenance
Predictive maintenance is maintenance triggered by a prediction of failure, based on measured equipment condition such as vibration, temperature, or noise.
Statistical process control (SPC)
Statistical process control (SPC) is the use of statistical methods, mainly control charts, to monitor a production process and detect when its variation goes beyond what its normal, stable behavior would produce.
Root cause analysis
Root cause analysis (RCA) is a structured method for finding the underlying cause of a failure or defect, so the corrective action prevents recurrence instead of only treating the symptom.
Soft sensor
A soft sensor, or virtual sensor, is software that estimates a quantity that is hard or costly to measure directly by calculating it from other measurements that are available in real time.