HomeBlogKnowledge Graphs + ML

Smarter fault diagnosis: merging knowledge graphs with machine learning

Accurate and explainable diagnosis — by grounding adaptive AI in a semantic map of the plant. Patent GB2619825A.

📅 23 Apr 2025 ✍️ Dr Max Cao ⏱️ 6 min read 🏷️ Fault Diagnosis

Key takeaways

  • Rule-based fault diagnosis is brittle; pure machine learning is accurate but hard to explain.
  • Golden Data combines a domain knowledge graph with adaptive ML for diagnosis that is both accurate and explainable.
  • A four-step method: semantic mapping → sensor integration → AI classification → feedback learning.
  • The approach is covered by patent GB2619825A and runs in the DPlus platform.

As industry digitalises, fault diagnosis is where a lot of value hides. Traditional rule-based systems are rigid and break when reality doesn't match the rulebook. Pure machine-learning models are flexible and accurate — but their "black box" nature makes engineers hesitate to act on them. The answer isn't one or the other; it's both, working together.

From raw data to intelligent diagnosis

By integrating semantic knowledge networks with adaptive AI models, manufacturers can dramatically increase reliability, reduce maintenance costs, and uncover hidden operational insights. The knowledge graph supplies the why — how equipment, components, faults and causes relate — while machine learning supplies the what's happening now from live sensor data.

Pipeline: sensor data and state vector feeding an ML model and a knowledge graph of symptom, fault, cause and asset
Knowledge graph + machine learning pipeline for fault diagnosis.

How the system works: four steps

1. Semantic mapping

A domain-specific knowledge graph captures equipment structure, failure modes and causal relationships — the engineering knowledge that usually lives in experts' heads.

2. Sensor integration

Real-time sensor data is connected to the graph, so each reading is interpreted in the context of the asset it belongs to.

3. AI classification

Machine-learning models classify fault types from the combined signal, predicting problems before they physically manifest.

4. Feedback learning

Outcomes feed back into the model, so diagnosis improves with every case — and the knowledge graph stays current.

GB2619825A
Patent covering the knowledge-graph + ML method
4-step
Semantic map → sensors → AI → feedback
Explainable
Diagnosis grounded in a semantic network

Why it matters

Grounding AI in a knowledge graph delivers two things black-box models can't: explainability — engineers can see why a fault was flagged — and adaptability — the system keeps learning. This is exactly the role our OODT semantic-mapping agent, Otto, plays inside DPlus: it builds the object network that every other diagnostic capability reasons over.

The takeaway

The future of industrial fault diagnosis isn't a smarter rulebook or a bigger model — it's a semantic foundation that makes machine learning trustworthy enough to act on.

MC

Dr Max Cao

Industrial AI & PHM specialist at Golden Data, focused on explainable diagnosis and knowledge-driven systems.

FAQ

Knowledge graphs + ML — questions

Why combine knowledge graphs with machine learning for fault diagnosis?

Knowledge graphs encode how equipment, faults and causes relate, while machine learning reads live sensor data. Combining them makes diagnosis both accurate and explainable — manufacturers can increase reliability, reduce maintenance costs and uncover hidden operational insights, instead of relying on rigid rule-based systems.

How does the four-step method work?

Semantic mapping builds a domain knowledge graph; sensor integration adds real-time data; AI classification predicts fault types; and feedback learning improves the model over time, enabling proactive detection before a fault physically manifests.

Is the diagnosis explainable?

Yes. Because predictions are grounded in a semantic knowledge network, the system can show why a fault was flagged, supporting trust and faster engineer action — a core advantage over black-box models.

Is the method patented?

Yes — the approach is covered by patent GB2619825A and is implemented in Golden Data's DPlus platform.

Take the next step

Make your diagnosis explainable

Related reading

More from the blog