💻 computer science

IndustryAssetEQA: A Neurosymbolic Operational Intelligence System for Embodied Question Answering in Industrial Asset Maintenance

IndustryAssetEQA is a neurosymbolic operational intelligence system that integrates episodic telemetry with a Failure Mode Effects Analysis Knowledge Graph (FMEA-KG) to provide grounded, verifiable, and counterfactual-capable embodied question answering for industrial asset maintenance.

Original authors: Chathurangi Shyalika, Dhaval Patel, Amit Sheth

Published 2026-04-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Chathurangi Shyalika, Dhaval Patel, Amit Sheth

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor in a massive, high-tech hospital. Instead of treating humans, you are treating giant, complex machines—like jet engines, massive water pumps, or factory robots.

These machines are constantly "talking" to you through thousands of digital sensors (their heart rate, blood pressure, and temperature). However, there is a problem: when a machine starts acting weird, the current AI assistants are like interns who have read a textbook but have never actually stepped into an operating room. If you ask them, "Why is this engine vibrating?" they might give you a beautiful, fluent answer like, "Vibration is often caused by mechanical wear," but they can't tell you which sensor saw the spike, when it happened, or what will actually happen if you replace a specific part. They are "smart" at talking, but "clueless" at doing.

IndustryAssetEQA is a new system designed to turn that "book-smart" intern into a "battle-hardened" specialist.

Here is how it works, using three simple metaphors:

1. The "Digital Detective" (The Fact Extractor & Episodic Store)

Instead of just looking at a messy pile of data, the system acts like a detective arriving at a crime scene. It doesn't just see "noise"; it gathers specific "evidence." It looks at a specific window of time (the "episode") and says, "At 3:00 PM, the pressure spiked by 20%, and the temperature was rising." It puts all this evidence into a neat, organized file cabinet (the Episodic Store) so it can always prove exactly where it got its information. No more guessing!

2. The "Master Encyclopedia" (The FMEA Knowledge Graph)

To make sure the AI doesn't just make things up, the researchers gave it a "Master Encyclopedia" of industrial wisdom called a Knowledge Graph. Think of this as a giant web of connections. It knows that Part A is connected to Sensor B, and if Sensor B goes crazy, it usually means Failure Mode C is happening. This prevents the AI from suggesting a "fix" that doesn't even exist for that specific machine. It grounds the AI's "imagination" in real-world engineering rules.

3. The "Flight Simulator" (The Risk Modeling & Causal Simulator)

This is the most important part. If you ask the AI, "What happens if I replace this seal right now?" a normal AI might say, "It will probably help." That’s not good enough when a jet engine is involved!

IndustryAssetEQA uses a "Flight Simulator." Before it answers, it runs a quick mathematical simulation: "If we change this variable, how does the risk percentage move?" It can then give you a precise, testable answer: "Replacing the seal will decrease the failure risk by 45%." It’s not just guessing; it’s simulating the future.

The Result: From "Chatty" to "Reliable"

The researchers tested this against standard AI (like basic versions of ChatGPT) and found a massive difference:

The Old Way: The AI was "overclaiming"—making confident-sounding guesses that were actually wrong or unsupported.
The IndustryAssetEQA Way: It became incredibly precise. It reduced "dangerous guesses" (overclaims) from 28% down to just 2%.

In short: This paper isn't just about making an AI that can talk about machines; it's about building an AI that can reason about them, prove its claims with evidence, and predict the consequences of our actions—making industrial maintenance safer and much smarter.

Technical Summary: IndustryAssetEQA

IndustryAssetEQA is a neurosymbolic operational intelligence system designed for Embodied Question Answering (EQA) in industrial asset maintenance. It addresses the critical reliability gap in current Large Language Model (LLM) deployments within safety-critical industrial environments.

1. The Problem: The "Reliability Gap" in Industrial LLMs

While LLMs provide fluent natural language interfaces, the authors identify three fundamental failures when they are used as standalone maintenance assistants:

Weak Grounding: LLMs often provide generic, "textbook" explanations that lack connection to the specific sensor telemetry of the asset in question.
Lack of Provenance: Answers rarely cite specific time windows, sensors, or maintenance records, making them unauditable for human operators.
Non-testable Reasoning: LLMs struggle with "what-if" (counterfactual) scenarios, offering qualitative suggestions (e.g., "fixing this might help") without quantitative or model-based risk assessments.

The paper argues that maintenance should be treated as an embodied decision problem—a loop of perception, reasoning, prediction, and action—rather than a simple text-generation task.

2. Methodology: A Neurosymbolic Architecture

The system integrates neural learning (LLMs) with symbolic reasoning (Knowledge Graphs and deterministic models) through a multi-component architecture:

FMEA Knowledge Graph (FMEA-KG): A domain-specific symbolic layer built from ISO-standard Failure Mode and Effects Analysis documentation. It maps assets to components, failure modes, sensors, and recommended maintenance actions.
Fact Extractor & Episodic Store: A deterministic pipeline that converts raw, multivariate time-series telemetry into structured "episode facts" (e.g., mean, standard deviation, and trends of specific sensors over a defined window). These are stored in a SQLite database to allow for precise, verifiable retrieval.
Risk Modeling & Causal Simulator: Instead of relying on the LLM to "guess" the effect of an action, the system uses a parametric multinomial logistic regression model. This simulator estimates the change in failure risk ( $\Delta r$ ) when a specific maintenance intervention is applied (e.g., "What if we replace the seal?").
EQA Builder & Prompt Builder: These components transform raw data and KG context into highly structured, task-specific prompts. They mandate that the LLM output a machine-parsable JSON containing the answer, reasoning, and explicit provenance.
Verifier and Safety Gate: A deterministic module that checks the LLM's output against the Episodic Store. If the LLM cites a sensor or a time window that does not exist in the actual data, the system flags it as an error.

3. Key Contributions

A New Framework: Formulating industrial maintenance QA as an embodied decision-making task.
IndustryAssetEQA System: A deployment-ready neurosymbolic architecture that combines telemetry, knowledge graphs, and risk simulators.
FMEA-KG: A structured, machine-interpretable knowledge graph for industrial failure semantics.
Benchmark Datasets: A collection of structured, time-situated episode facts spanning four industrial asset types (rotating machinery, turbofans, hydraulics, and cyber-physical systems).

4. Results and Evaluation

The system was evaluated against LLM-only baselines (GPT-4o-mini and Claude Sonnet 4) across five QA task types (Descriptive, Temporal, Diagnostic, Counterfactual, and Action-Oriented).

Reliability Gains: IndustryAssetEQA significantly outperformed LLM-only baselines. It improved structural validity by up to +0.51 and counterfactual accuracy by up to +0.47.
Reduction in Overclaims: Expert human evaluation showed that "severe overclaims" (unsupported or unsafe statements) dropped from 28% in LLM-only models to just 2% in the full IndustryAssetEQA system.
Ablation Insights:
- Removing the Risk Simulator crippled counterfactual reasoning.
- Removing Provenance Enforcement caused the largest collapse in deployable reliability.
- The FMEA-KG was essential for semantic alignment and explanation faithfulness.

5. Significance

The significance of this work lies in its move away from "black-box" AI toward verifiable operational intelligence. By anchoring LLM reasoning in physical telemetry (perception), domain knowledge (reasoning), and probabilistic risk models (prediction), IndustryAssetEQA provides a blueprint for AI assistants that can be trusted in high-stakes, safety-critical industrial settings. It transforms the AI from a "chatbot" into a "decision-support agent" capable of providing auditable, evidence-backed, and risk-aware maintenance recommendations.