Conversational AI for Reliable Insights from Industrial Telemetry (with Alfa Laval)
| Title | Conversational AI for Reliable Insights from Industrial Telemetry (with Alfa Laval) |
|---|---|
| Summary | collaborate with Alfa Laval (a leading national and global company); Conversational AI for industrial telemetry, combining language models with numerical data and documentation to deliver reliable, explainable insights on machine status and performance. |
| Keywords | |
| TimeFrame | Spring 2025 |
| References | |
| Prerequisites | Good knowledge of machine learning |
| Author | |
| Supervisor | Mahmoud Rahat, Saeed Gholami Shahbandi |
| Level | Master |
| Status | Open |
Conversational AI for Reliable Insights from Industrial Telemetry
Opportunity to collaborate with Alfa Laval, a leading national and global company.
Introduction
This project explores conversational AI for industrial telemetry: enabling a system to answer questions about recent operations, KPIs, diagnostics, prognostics, and root cause analysis across machine --> vessel --> fleet. The work fuses language models with structured numerical time series and authoritative technical documents to produce grounded, citeable answers rather than freehand text. Emphasis is on numerical reliability, clear provenance, and explainability in real-world settings.
Using real industrial datasets, in collaboration with Alfa Laval, the project investigates how to format telemetry for interpretability (unit-aware schemas, anomaly-preserving rollups), how to provide essential context (temporal windows, baselines, thresholds), and how to ground answers (retrieval, secure tool-use, validators) to reduce hallucination and improve factuality. A shared reference approach and evaluation protocol will guide several focused thesis projects that contribute complementary components within this overall scope.
Prospective thesis project topics
Tool-Augmented Telemetry Reasoning for Conversational Interfaces
Objective: Ensure numerically correct, citable answers by computing first (secure tools), validating (units/ranges/baselines), then citing results in chat.
Scope: Planner vs. retrieval-first policies; telemetry stats (windowed means, anomalies, CIs); RBAC; evidence citations; optional RAG for limits/specs.
Expected outcome: Reference agent + validator toolkit; policy benchmark for factuality vs. latency.
RQs: Which tool policy best balances factuality and latency? Which validators (units/thresholds/temporal baselines) cut numeric errors most?
Schema- & Context-Optimized Telemetry Formats for LLM Numeric Reasoning
Objective: Improve interpretability via unit-aware “stat-cards” and explicit temporal context (now/Δ/rolling baseline) with lightweight prompting or minimal tuning.
Scope: Schema variants (JSONL, compact tables, key-value with units/limits); anomaly-preserving rollups; prompt templates vs. small instruction-tunes.
Expected outcome: Schema guidelines + converters + prompt library; ablations of context components vs. accuracy/latency.
RQs: Which schema+context combo maximizes numeric correctness per token? How much do spec guardrails and temporal packaging reduce errors?
