AI Glossary / Governance, ethics and safety / Intermediate

Data fusion

Definition

Data fusion combines information from multiple sources to form a joint estimate, representation, or assessment. The sources may provide overlapping measurements or complementary evidence about the same situation. Fusion aims to produce information that is more useful than isolated inputs, but improvement is not automatic: source quality, alignment, dependence, and the combining method all affect the result.

What it means

A temperature sensor and a camera may describe different aspects of a manufacturing process. Two sensors may instead estimate the same quantity with different levels of uncertainty. Data fusion covers both kinds of situation, although their methods differ. The term is broader than copying files into one location. A shared repository makes data accessible; fusion uses relationships among sources to form a combined interpretation. Boundaries with data integration vary across fields, so an explanation should identify what information is combined and what new output is produced.

How it works

Begin by defining the entity, event, or state being estimated. Establish whether records refer to the same object and align units, timestamps, and coordinate systems. Fusion can occur at the raw-data level, after extracting features, or after separate models have produced decisions or scores. A statistical method may combine measurements according to uncertainty, while a learned model may combine features from several sources. Preserve provenance so that the result can be traced back to its inputs. When sources disagree, the procedure needs an explicit response: retain uncertainty, flag a conflict, or apply a justified reliability model. Silently averaging incompatible measurements can hide a problem rather than resolve it.

Three practical examples

• A factory combines vibration and temperature measurements to assess equipment condition, checking that both refer to the same machine and time interval.
• A mapping system aligns overlapping sensor observations to estimate an object's position more precisely under a documented measurement model.
• A document workflow combines extracted text with page-layout features to identify a table entry. Each source contributes different evidence, and the system should retain enough information to investigate a mistaken association.

Think of it this way

Think of assembling a puzzle from several photographs of the same scene. One photograph may show a detail another misses, but the images must be aligned before they can be combined meaningfully. The analogy has limits: data sources can share errors, use different definitions, or describe different moments. More pieces do not necessarily mean a truer picture if the pieces are duplicates or have been matched to the wrong scene.

Worked example

Suppose two independent, unbiased sensors measure the same stable temperature. Sensor A reports 20 degrees with error variance 1, and sensor B reports 24 degrees with error variance 4, using the same squared units. Under this simplified model, inverse-variance weights are 1 and 1/4. Normalizing gives weights 0.8 and 0.2, so the combined estimate is 0.8 times 20 plus 0.2 times 24 = 20.8 degrees. Its error variance under the assumptions is 1/(1 + 1/4) = 0.8. This calculation does not justify combining arbitrary readings. If sensor B measures a different location, the quantity is changing rapidly, or both sensors share a calibration bias, the assumptions fail. Two copies of the same sensor reading must not be treated as independent evidence that reduces uncertainty.

Illustrative case study

Fictional case: Hawthorn Manufacturing combines camera observations with process-temperature readings to investigate defects. Initial results associate unusual images with the wrong temperature spikes because the device clocks are offset. After correcting the time alignment, the team compares the fused system with each source on its own and retains cases where the sources disagree. It also records which sensors contributed to each assessment. The case shows that careful registration and traceability can matter as much as the choice of an AI model; it does not claim a measured improvement for a real company.

Common misconceptions and mistakes

More sources do not necessarily provide more independent information. Several feeds may reproduce the same original record or inherit the same error. A combined score is also not automatically a calibrated probability. Entity matching mistakes can fuse records about different people, products, or events, creating an apparently coherent but false account. Missing data should not be silently replaced with a reassuring default. Finally, fusion can reveal information that was not obvious in any source alone, so the suitability of the combined use needs consideration rather than evaluating each source only in isolation.

Business use

Define the purpose of the combined dataset or estimate and identify who is responsible for its quality. Document source ownership, permitted uses, retention, and access controls as appropriate to the project. Validate alignment and matching before interpreting model performance. Test the effect of losing or corrupting one source, and compare the fused approach with simpler single-source baselines. Keep uncertainty and conflicting evidence visible to decision makers. When information concerns people, assess whether linking sources changes the sensitivity or implications of the result. This is a governance design question as well as a technical integration task.

People also ask

Is data fusion always an AI technique?

No. It can use direct rules, statistical estimation, signal processing, or machine learning.

Can it combine different data types?

Yes. Text, images, sensor readings, and other representations can be combined through an appropriate method.

Does fusion remove disagreement between sources?

It may resolve some discrepancies under a justified model, but unresolved disagreement should sometimes remain visible rather than being forced into one answer.

Check your understanding

Three reports agree, but all copied the same original measurement. Should a fusion system count them as three independent confirmations?

Answer: No. Their shared origin matters. Treating duplicates as independent can create unjustified confidence in the combined result.

Related terms

• Multimodal Learning
• sensor fusion
• data integration
• entity resolution
• data provenance
• uncertainty quantification
• feature fusion
• decision fusion

Learn next

Choose two sources that describe the same operational event. Write down how their identifiers, times, units, and errors differ. Specify a rule for disagreement and a test using deliberately mismatched records. This reveals whether the proposed fusion actually combines compatible evidence.

More in Governance, ethics and safety

• Dark pattern - A dark pattern is a design practice that misleads, pressures or obstructs people in ways that undermine informed choice.
• Data control - Data control is the practical ability to set and enforce conditions on how data is collected, accessed, changed, used, shared and retained.
• Data dredging - Data dredging is searching through many analyses or patterns and presenting selected favorable findings without adequately accounting for the search.
• Data Generation Bias - Data generation bias is a distortion introduced when synthetic or redundant examples are added to a dataset in ways that misrepresent the intended...
• Data point - A data point is one unit of information considered in an analysis or machine-learning task.
• Data poisoning - Data poisoning is deliberate interference with data used by an AI system in order to cause unwanted behavior.

Sources

NIST, Data Integration, Management and Fusion for AM Industrialization: https://www.nist.gov/programs-projects/data-integration-management-and-fusion-am-industrialization Yang and colleagues, In-Process Data Fusion for Process Monitoring and Control of Metal Additive Manufacturing: https://www.nist.gov/publications/process-data-fusion-process-monitoring-and-control-metal-additive-manufacturing

Previous: Data-flow diagram Next: Data generating process