AeroGuard TSLM

Project Overview

Built during the European Hackathon League (EHL Hackathon Zurich 2026) under the "Give AI a Sense of Time" challenge hosted by Aionic Labs and the ETH Agentic Systems Lab (ASL), AeroGuard TSLM bridges continuous multivariate jet engine telemetry directly into the embedding space of a causal language model (SmolLM-135M-Instruct) via Multimodal Prefix Token Fusion.

In a unified forward pass, AeroGuard simultaneously forecasts numerical Remaining Useful Life (RUL) with sub-cycle precision and synthesizes actionable, physically grounded Chain-of-Thought (CoT) aerothermal engineering diagnostics for fleet reliability engineers and Maintenance, Repair, and Overhaul (MRO) crews.

Problem & Industrial Motivation

In commercial aviation, unscheduled engine removals (Aircraft on Ground / AOG) cost operators upwards of $150,000+ per incident due to cascade flight delays, passenger accommodations, and emergency charters. Traditional maintenance approaches face critical bottlenecks:

Turbofan Engine Station Mapping & Telemetry

Jet engine wear is strictly governed by coupled thermodynamics across consecutive gas-path stations. AeroGuard monitors 14 dynamic telemetry channels across primary engine components:

Station-by-station telemetry mapping: FAN (Nf, NRf), LPC (T24), HPC (T30, P30, Ps30, Nc), Combustor, HPT (T50), LPT, and Exhaust Nozzle.

Core Technical Innovations

• Multimodal Prefix Token Fusion: Continuous sensor waveforms (window length = 30 flight cycles) are projected through a 1D convolutional patch encoder directly into LLM token embeddings, avoiding discrete quantization errors and table serialization limits.
• Multi-Task Joint Loss Objective: Optimizes autoregressive language modeling cross-entropy jointly with an auxiliary scalar RUL regression head:
  L_total = L_LM + 0.01 × L_RUL.
• 4-Stage Aerothermal Chain-of-Thought (CoT): Synthesizes transparent diagnostics across: (1) Sensor Drift Observation, (2) Aerothermal Root Cause Analysis, (3) Fleet Health & Criticality Bounding, and (4) Line-Replaceable Unit (LRU) MRO Action Plan.
• Strict Zero-Leakage Data Partitioning: Engines partitioned strictly by physical unit ID on NASA C-MAPSS FD001 (Train: 1–70, Val: 71–80, Test: 81–100); normalization statistics computed exclusively on training units.

Benchmark Results on Held-Out Test Engines

Rigorous evaluation on held-out test engines (Units 81–100) demonstrated decisive performance advantages against state-of-the-art baselines:

• AeroGuard TSLM (Ours): Achieved 7.94 RMSE on held-out test engines — an 85% error reduction compared to Amazon Chronos T5 (53.93 RMSE).
• NASA Asymmetric Scoring Metric: Lowered asymmetric late-prediction penalty score from 17.22M (Chronos) down to 2.62M.
• Reasoning Reliability: 100% valid, physically grounded diagnostic response rate with complete elimination of tabular hallucinations.
• Interactive Mission Control Dashboard: Streamlit web application with 3D turbofan cross-sections, synchronized Plotly telemetry inspection, and automated ETOPS flight clearance dispatch.

Hackathon Presentation Highlights

Key presentation slides delivered at the EHL Zurich 2026 hackathon finale:

Problem Statement & Cost of Unscheduled Removals

Held-out Benchmark Results vs. Amazon Chronos

Team & Hackathon Details