Industrial Knowledge Learning and Deployment Research for Utilities
Industrial AI must remain useful as equipment, data, and operating conditions change. This research program examines how utility organizations can organize operational knowledge, learn from limited evidence, control model updates, and reuse validated capabilities across substations, feeders, transformers, switchgear, protection systems, and other grid assets.
Explore research areas ↘Building a Connected Industrial Knowledge Layer
Utilities operate with a mixture of engineering drawings, asset registers, supervisory control and data acquisition (SCADA) records, geographic information system (GIS) data, computerized maintenance management system (CMMS) histories, inspection reports, alarm records, work orders, operating procedures, and technician knowledge.
These sources often use different identifiers, naming conventions, formats, and levels of completeness. A useful industrial knowledge layer must connect those records without obscuring uncertainty or creating unsupported relationships.
The research also addresses a practical deployment problem: a model or decision workflow that performs well at one site may encounter different equipment configurations, sensor behavior, maintenance practices, or operating conditions elsewhere. Research therefore considers controlled adaptation, expert feedback, monitoring, transfer evaluation, and rollback rather than assuming that a trained model can be deployed unchanged.
Key Research Questions
The research program examines how utility knowledge can be represented, evaluated, improved, and deployed across different sites while maintaining traceability and appropriate governance.
Consistent Utility Data Semantics
How can utility asset records, equipment documentation, inspection history, SCADA observations, maintenance records, and operator knowledge be represented with consistent semantics?
Semantic Interoperability
Where can Asset Administration Shells and related semantic models improve interoperability among equipment records, engineering information, and AI workflows?
Learning from Limited Evidence
How can learning methods use limited labels and structured expert feedback without treating expert judgment as ground truth in every situation?
Transfer to Unfamiliar Sites
How should models be evaluated when transferred to an unfamiliar substation, transformer population, feeder configuration, or operating condition?
Monitoring and Controlled Updates
What monitoring signals should trigger review, adaptation, retraining, release, or rollback when data distributions or operating conditions change?
Operator and Teleoperation Learning
How can operator interventions, maintenance decisions, and teleoperation records be captured as learning evidence while preserving appropriate data rights and governance?
Cross-Site and Federated Learning
When can cross-site or federated learning reduce data movement while retaining clear ownership, access, privacy, and validation boundaries?
Economic and Deployment Evaluation
How should complete deployment costs and measured operational benefits be evaluated so that pilot economics are not confused with guaranteed future returns?
Industrial Knowledge, Learning and Deployment
The research areas connect utility information management with learning methods, operational feedback, monitoring, and controlled deployment.
Industrial Knowledge Integration and Semantic Interoperability
A transformer may have an engineering identifier in one system, a maintenance identifier in another, a GIS representation, SCADA tags, inspection records, and manufacturer documentation stored elsewhere.
Research can examine common semantic structures that preserve these relationships while recording source, timestamp, unit, quality, and confidence.
Asset Administration Shells can be investigated as one approach to representing structured information about industrial assets and their associated data. The objective is to reduce mapping effort, support reusable interfaces, and make the context behind AI inputs easier to inspect.
Learning from Limited Labels and Expert Feedback
Many utility workflows do not produce large, consistently labeled datasets. Failure events may be relatively rare, inspection findings may be documented differently by different teams, and operating records may contain gaps.
Research can examine learning methods that make better use of limited labels, engineering rules, inspection findings, and structured expert review.
Expert feedback should be separated from independent test evidence. Engineer assessments can inform later model development, but should not automatically be treated as independent proof of performance.
Model Monitoring, Adaptation and Reviewed Release
Utility data can change when sensors are replaced, firmware is updated, equipment is refurbished, operating practices change, new assets are commissioned, or a site begins operating under different load conditions.
Monitoring should examine input quality, data distribution, output behavior, uncertainty, and operational usefulness.
Candidate model updates should be evaluated against a fixed baseline, reviewed for regressions, versioned, and released only after defined acceptance checks. A rollback path should remain available.
Operator and Teleoperation Data as Learning Evidence
Operators, control-room staff, field technicians, protection engineers, and maintenance teams routinely make decisions that are valuable for understanding how utility systems behave.
Research can examine how interventions, overrides, inspection outcomes, maintenance actions, and teleoperation records can be captured as structured learning evidence.
Useful records include the observed state, action taken, reason for intervention, subsequent outcome, and level of human authority.
Cross-Site and Federated Learning
Utilities may need to learn across sites without centralizing all operational data. Federated learning is one candidate method, but it should be evaluated as a system design rather than treated as a privacy guarantee.
Research should define data rights, participating assets, model-update controls, validation procedures, and what information can leave each site.
Evaluation should consider task performance, communication and integration effort, privacy exposure, model regressions, review workload, and site-specific differences.
Economic Evaluation and Deployment Readiness
A technically successful model is not automatically an economically useful deployment. Utility economics should account for data preparation, equipment and edge hardware, integration, software, model monitoring, maintenance, training, cybersecurity controls, and recurring support.
Benefits should be measured against a defined baseline and time period.
Research should distinguish predicted value from measured value. Pilot estimates should remain tied to their assumptions and operating conditions.
Connecting Utility Asset Records and Operational Knowledge
A semantic knowledge layer can connect information created for different operational purposes while retaining the context and quality of each source.
Asset Identity and Relationships
Connect engineering identifiers, maintenance identifiers, GIS representations, SCADA tags, inspection records, and manufacturer documentation to the relevant equipment.
Traceable Information Context
Preserve source, timestamp, measurement unit, data quality, and confidence so that relationships and AI inputs can be inspected.
Reusable Semantic Interfaces
Investigate Asset Administration Shells and related semantic models as ways to reduce mapping effort and support reusable interfaces.
Existing Utility Systems
The objective is not to replace existing utility information systems, but to improve connections among records while retaining their original meaning.
From Learning Evidence to Reviewed Deployment
Industrial deployment requires more than initial model accuracy. A controlled process should account for changing conditions, expert review, release acceptance, and rollback.
Collect and Contextualize
Connect utility data, equipment context, inspection findings, operating records, and expert knowledge with clear identifiers and data-quality information.
Learn and Evaluate
Use suitable learning methods and structured expert feedback. Evaluate performance on familiar and unfamiliar equipment or operating conditions.
Monitor and Review
Monitor input quality, distribution changes, output behavior, uncertainty, and operational usefulness. Review candidate changes before release.
Release and Roll Back
Version approved changes, apply defined acceptance checks, and retain a rollback path when a new version performs worse than expected.
Establish the Baseline
Represent a local workflow using existing mappings, rules, and site-specific analysis.
Introduce the Candidate Approach
Introduce shared interfaces, a reviewed semantic mapping layer, or controlled model adaptation.
Evaluate Across Sites
Compare mapping effort, association quality, task performance, expert review time, and regressions after transfer to another site or equipment group.
Document Evidence Status
Record which elements are simulated, derived from authorized operational data, reviewed by subject matter experts, and independently evaluated.
Evaluation Measures
Evaluation should measure both technical performance and the operational effort required to maintain a useful deployment.
Asset-record mapping effort and mapping error rate.
Identity and semantic association errors across equipment records.
Task performance on familiar and unfamiliar sites or equipment.
Expert review hours and the proportion of cases requiring manual correction.
Regression frequency and severity after model or knowledge-base updates.
Rollback time and the completeness of the rollback process.
Integration effort required to connect source systems and reusable interfaces.
Privacy or data-exposure measures appropriate to the selected cross-site learning design.
Difference between predicted operational value and measured value after controlled evaluation.
Complete deployment cost, including hardware, integration, software, maintenance, training, and recurring support.
Learning from People, Operations and Distributed Sites
Operational knowledge can come from human decisions and from multiple utility locations. Both require clear context, access boundaries, and evaluation.
Operator and Teleoperation Data
Operators, control-room staff, field technicians, protection engineers, and maintenance teams make decisions that can help explain how utility systems behave. Research can examine how interventions, overrides, inspection outcomes, maintenance actions, and teleoperation records can be captured as structured learning evidence.
Useful context includes the observed state, action taken, reason for intervention, subsequent outcome, and level of human authority. These records require consistent capture, clear definitions, and appropriate access controls.
Cross-Site and Federated Learning
Federated learning is a candidate method for learning across sites without centralizing all operational data. It should be evaluated as a system design rather than treated as a privacy guarantee.
A cross-site study should compare local-only learning with a shared or federated approach under controlled conditions, considering task performance, communication and integration effort, privacy exposure, model regressions, review workload, and site-specific behavior.
Expert Feedback and Independent Evidence
Structured expert feedback can inform later model development, but the same record should not automatically be treated as independent evidence of model performance.
Evaluation should test whether learning transfers across equipment types, maintenance crews, substations, and operating conditions.
Economic Evaluation and Deployment Readiness
Utility economics should account for the full lifecycle of a solution, including data preparation, equipment and edge hardware, operational-system integration, software, model monitoring, maintenance, training, cybersecurity controls, and recurring support.
Benefits should be measured against a defined baseline and time period. Predicted value and measured value should remain distinct.
Candidate Research Methods
Depending on the research question, methods may include semantic data modeling, record linkage, knowledge graphs, supervised or semi-supervised learning, anomaly detection, transfer learning, model monitoring, controlled adaptation, federated learning, and structured human review.
Utility Data Sources
Relevant sources may include asset registers, GIS records, SCADA measurements, protection and control records, inspection reports, maintenance histories, work orders, equipment manuals, alarm logs, operating procedures, and expert annotations.
Data Quality and Governance
Each source should have defined identifiers, timestamps, units, quality indicators, ownership, access rights, and retention requirements.
Synthetic and Simulated Data
Where synthetic or simulated data are used, retain the scenario specification, assumptions, data dictionary, generation code, fixed configuration, random seeds, evaluation outputs, and limitations. Synthetic results should not be presented as field measurements or customer results.
Research Status and Evidence
Topic inclusion describes a proposed area of investigation. It does not establish a completed product, deployed utility system, customer result, or separate research team.
Make the evidence level clear.
Research evidence should distinguish proposed investigation, simulation using synthetic data, controlled bench or hardware testing, customer pilot evidence, and operational customer results.
Specific demonstrations and findings should be published only with an appropriate evidence label and after technical review. Claims should be updated when pilot evidence changes assumptions, evaluation results, or deployment boundaries.
A simulation-based demonstration should not be presented as evidence of an actual customer deployment.
Limitations and Practical Considerations
Knowledge integration and industrial learning can improve utility workflows, but their practical limits must remain visible during evaluation and deployment.
Semantic interoperability does not remove the need to resolve conflicting identifiers, incomplete records, or unclear asset ownership.
Learning from expert feedback can introduce bias or inconsistent labeling if review procedures are not defined.
Transfer performance can decline when equipment, operating conditions, instrumentation, or maintenance practices differ from the development environment.
Model monitoring can detect changes but does not by itself determine whether a change is operationally safe or economically important.
Federated learning does not guarantee privacy and still requires governance of model updates, metadata, access, and participating data sources.
Pilot economics do not establish future returns without measurement under the relevant operating conditions.
