Entity Resolution

A rule-based approach that establishes identity by applying exact or fuzzy comparison logic across a defined set of attributes. In manufacturing, this typically involves matching on composite keys like manufacturer + serial number or equipment tag ID.

• Exact Matching: Requires strict equality on a unique identifier (e.g., asset_id == 'PUMP-23')
• Fuzzy Matching: Uses algorithms like Levenshtein distance or phonetic encoding to tolerate typographical errors in human-entered maintenance logs
• Rule Cascades: Engineers define a sequence of match rules executed in priority order, such as first attempting a serial number match, then falling back to a location + model number combination

This method is highly transparent and auditable but struggles with the high variability and incompleteness of legacy industrial data.

A statistical framework, formalized by Fellegi and Sunter, that calculates the probability that two records refer to the same entity based on the agreement or disagreement of multiple fields. Each attribute is assigned a weight reflecting its discriminatory power.

• M-Probability: The likelihood that a field matches given the records are a true match
• U-Probability: The likelihood of a coincidental match by random chance
• Composite Weight: A total score is calculated by summing the log-likelihood ratios of all field comparisons, with a decision threshold classifying pairs as matches, non-matches, or requiring clerical review

This technique excels at resolving assets across siloed systems like a CMMS and a SCADA historian where no single unique key exists.

A computational optimization strategy that avoids the intractable O(n²) complexity of comparing every record against every other record. Blocking partitions the dataset into mutually exclusive buckets using a blocking key, and comparisons are only performed within each bucket.

• Sorted Neighborhood: Records are sorted on a key (e.g., asset name) and a sliding window of fixed size is moved over the sorted list, comparing only records within the window
• Canopy Clustering: A cheap, approximate distance metric is used to rapidly group records into overlapping canopies before applying an expensive, precise comparison
• Manufacturing Application: Blocking on plant location or equipment class drastically reduces the search space when linking millions of sensor tags to a master asset register

The application of supervised and active learning models to classify record pairs as matches or non-matches, moving beyond hand-tuned rules. A training set of labeled pairs is used to teach a model the complex, non-linear interaction of field similarities.

• Feature Engineering: Input features include Jaccard similarity, TF-IDF cosine distance, and numeric difference ratios for each attribute
• Classifier Models: Algorithms like Random Forest, Gradient Boosting, or Siamese Neural Networks learn the optimal weighting and interaction of these features
• Active Learning: The model identifies the most uncertain record pairs and requests human labels, dramatically reducing the manual labeling burden required to achieve high accuracy on noisy factory-floor data

A technique that leverages the connected structure of data to resolve identity, moving beyond pairwise attribute comparison. Entities are modeled as nodes, and their relationships—such as sharing a common part, being maintained by the same technician, or reporting to the same gateway—provide strong disambiguating signals.

• Collective Resolution: The decision to merge two asset nodes propagates new evidence through the graph, influencing adjacent match decisions in an iterative process
• Community Detection: Algorithms like Louvain or Label Propagation can cluster nodes that are densely interconnected, revealing hidden groupings of records that all refer to the same physical machine
• Contextual Evidence: A vibration sensor and a temperature sensor reporting from the same IP address and physical location are strong graph-based indicators they monitor the same asset, even if their naming conventions differ completely

The final stage of entity resolution where matched records are merged into a single, authoritative, and non-redundant representation of the real-world asset. This is not a simple union; it requires a survivorship strategy to resolve conflicting values.

• Attribute-Level Rules: Define which source system is authoritative for each field (e.g., the ERP is the source of truth for purchase date, while the CMMS is authoritative for maintenance history)
• Conflict Resolution Functions: Strategies include LONGEST_VALUE, MOST_RECENT, MOST_FREQUENT, or a custom voting algorithm
• Provenance Tracking: The golden record retains a pointer to every source record that contributed to it, maintaining full lineage for auditability and allowing the reconstruction of the merge decision at any point in time

A formal, explicit specification of a shared conceptualization that defines the types, properties, and interrelationships of entities within a manufacturing domain. Entity resolution relies on ontologies to define equivalence semantics—without a formal model specifying that hasSerialNumber is a unique identifier or that isLocatedAt is a transitive property, the system cannot logically determine if two records refer to the same physical asset. Ontologies provide the vocabulary and constraints that make disambiguation deterministic.

The fundamental data structure of a knowledge graph consisting of a subject-predicate-object statement. After entity resolution links disparate records into a single golden record, the merged entity is represented as a set of consolidated triples. For example, resolving two records for 'Pump-23' eliminates duplicate triples and reconciles conflicting ones, producing a canonical set like:

• <Pump-23> <hasFailureMode> <BearingFatigue>
• <Pump-23> <installedOn> <2023-06-15>
• <Pump-23> <locatedIn> <Line-4>

A W3C standard for validating RDF graphs against a set of conditions. SHACL shapes enforce data quality post-resolution by verifying that merged golden records conform to expected patterns. For instance, a SHACL constraint can assert that a resolved asset entity must have exactly one unique serial number and at least one location assignment. If entity resolution produces a record with conflicting serial numbers that cannot be reconciled, the SHACL validator flags it for human review, ensuring the knowledge graph remains trustworthy.

A specialized knowledge graph that captures the complete lineage of a data point, recording its origin, all transformations, and the agents involved. Entity resolution decisions must be auditable—a provenance graph tracks exactly which source records were merged, which matching algorithm was used, and the confidence score of each linkage. This is critical in regulated manufacturing environments where an incorrect merge could lead to a defective batch. The provenance graph enables rollback and forensic analysis of resolution decisions.

An Industry 4.0 standard for a digital representation of a manufacturing asset that provides a standardized, interoperable manifest of its properties, capabilities, and sub-models. Entity resolution is the process that populates a unified AAS from fragmented data sources. When an ERP system, a PLC historian, and a maintenance log all contain partial information about the same physical motor, entity resolution merges these fragments into a single, comprehensive AAS that serves as the asset's authoritative digital twin node.

A query execution strategy that decomposes a single semantic query across multiple distributed, autonomous graph databases and aggregates the results. Entity resolution often operates in a federated architecture where raw data remains in source systems due to security or sovereignty constraints. A federated query engine executes matching algorithms across these silos without centralizing data, using techniques like privacy-preserving record linkage to identify matching entities while keeping sensitive operational data behind each factory's firewall.