smartnanotubes/snt-fire-enose-5class

SNT Fire/Smoke 5-Class Classifier

Part of SmartNanotubes' work on machine olfaction — the AI layer for chemical sensing. SNT builds carbon-nanotube (CNT) electronic-nose hardware and the machine-learning models that turn raw sensor response into decisions. This page documents the first publicly validated component of that stack: a fire/smoke hazard classifier, reported with full cross-device and cross-site rigor.

Explore this work:

• 🎬 Demo video: watch the e-nose in action
• 🧪 Sample data: smartnanotubes/snt-fire-enose-sample — real recordings you can download and tinker with
• ▶️ Try it live: smartnanotubes/snt-fire-enose-demo — classify a recording in your browser

Machine olfaction is an emerging field with very few organizations combining a real sensing platform, deployed hardware, and honestly benchmarked models. SNT is one of them. Rather than publishing an unverifiable "universal nose" claim, we are publishing specific models with measured generalization — starting here — and will extend this presence as further components are validated.

This card documents the model's intended use, its validation methodology, and its measured generalization performance across different devices and different physical sites — the two conditions that matter most for real-world deployment. Model weights are not distributed here; see How to Get Started.

Table of Contents

• Model Details
• Uses
• Bias, Risks, and Limitations
• Training Details
• Evaluation
• Citation & Lineage
• Model Card Authors & Contact
• How to Get Started with the Model

Model Details

Model Description

The model performs multi-class classification of air-quality events from the time-series output of a CNT chemiresistive gas-sensor array. Given a window of sensor response, it assigns one of five classes:

• Normal air (background / no event)
• Cigarette smoke
• Cooking / frying aerosols
• Smoldering electrical cable
• Smoldering fabric / tissue

The classifier is designed for early hazard detection in monitored indoor environments, with an explicit design priority on low false-alarm rates (a system that alarms on ordinary cooking is not useful).

• Developed by: SmartNanotubes Technologies (SNT)
• Model type: Supervised multi-class classifier over engineered sensor time-series features
• Sensor modality: 16-channel CNT chemiresistive array, 1 Hz sampling
• License: Proprietary (see LICENSE)

Model Sources

• Developer: https://www.smart-nanotubes.com/
• Platform lineage: see Citation & Lineage

Uses

Direct Use

Early detection and classification of fire/smoke-related air events in monitored indoor environments (e.g. residential, commercial, and some industrial spaces) using an SNT e-nose. The model is intended to run locally on the e-nose device as an edge AI, outputting real-time classification of smells in the air.

Out-of-Scope Use

This model is not suitable for, and must not be relied upon for:

• Certified life-safety fire detection. This is a research/demo model, not a certified fire-alarm device under any national or international safety standard.
• Gas concentration measurements. The model classifies event type; it does not report quantitative concentration.
• Open-set gas or odor identification. It recognizes only the five trained classes; it is not a general-purpose "identify any smell" system.

Bias, Risks, and Limitations

• Cross-site generalization gap. In-domain performance (same devices and sites as training) can be higher than performance on entirely new sites. The honest deployment number is the leave-one-site-out result in Evaluation, not the in-domain number.
• One class is weaker than the others. Under the strictest cross-site test and using the simplified feature set, one of the five classes is considerably harder to separate from the others; performance on that class should not be assumed adequate for safety-critical use without site-specific validation.
• Environmental sensitivity. Ambient temperature and humidity influence chemiresistive response and can affect classification. Environmental compensation and per-installation calibration are not included in the simplified feature set.
• Long-term drift. As with all chemiresistive gas sensors, sensor response can drift over time (years). Periodic re-calibration / re-anchoring is not included here, as all presented data were acquired within a few months' time window.
• Not a substitute for certified detectors. Any safety-critical deployment must retain certified detection equipment.

Training Details

Training Data

The model was trained on controlled exposure recordings collected with SNT e-nose devices: normal-air baselines plus controlled generation of each fire/smoke class (cigarette smoke, cooking aerosols, smoldering cable, smoldering fabric). Data was gathered across 100+ sensor chips and spanning 20+ independent physical sites. The sensor chips were not pre-selected and reflected the full chip-to-chip variability after production. The data do not contain any sensor-variability compensation. This is done to provide the whole possible spread of sensor responses over different chips, sites, and sessions to support generalization testing.

Dataset specifics (exact composition, provenance, and preprocessing) are proprietary and not released.

How readings were taken

The data were collected using SNT's SmellBoard devices. Each device contains four 16-channel smell-sensor chips, one temperature sensor, and one humidity sensor. Every second the device outputs a string in ASCII format with 66 values: 4 groups of 16 resistance values in Ohm, temperature in °C, and relative humidity in %. Therefore, at least 4 sensor chips (1 SmellBoard device) see the same session at the same time and site, and corresponding environmental data is always provided.

Close-up of the CNT detector chip:

The sensor module — four 16-channel CNT sub-arrays on the plug-in board:

• ▶️ Watch the demo video — see the e-nose capturing and classifying an event in real time.

Training Procedure

At a high level, raw multi-channel sensor streams are converted into engineered time-series features that capture both signal amplitude and the adsorption/desorption kinetics of the sensor response, and a gradient-boosted tree ensemble is trained on these features with class-balanced objectives to counter the strong class imbalance (normal-air dominates).

Specific feature definitions, hyperparameters, and the detector-chip composition details are proprietary and not disclosed in this card.

Compute: Trains and runs on commodity CPU hardware; inference is lightweight (sub-second per window). No GPU required for inference.

Evaluation

Testing Data, Factors & Metrics

• Testing data: Held-out recordings from the proprietary multi-site dataset.
• Factors: Two generalization axes are evaluated explicitly — the device a sample came from, and the site it was recorded at.
• Metric: Macro-F1 is the headline metric. Because normal-air dominates the data, plain accuracy is misleading; macro-F1 weights every class equally and is the honest measure of multi-class performance under imbalance.

Three evaluation protocols are reported, from most optimistic to most honest:

Protocol	What is held out	What it measures
In-domain (random split)	Random samples	Best case; same devices and sites in train and test
Leave-One-Device-Out (LODO)	All data from a device	Generalization to a new device at known sites
Leave-One-Site-Out (LOSO)	All data from a site	Generalization to a new site — the real-world deployment case

Why this matters. Randomly splitting sensor windows lets data from the same recording session appear in both train and test, which inflates scores toward ~1.0 and is not a valid measure of generalization. All results below use event-grouped splits (whole recordings held out together), and the site-level (LOSO) number is the one that reflects deploying to a genuinely new environment.

Results

Headline macro-F1:

Protocol	Macro-F1
In-domain (random split)	≈ 0.88
Leave-One-Device-Out	≈ 0.83
Leave-One-Site-Out	≈ 0.73

Per-class F1 (class labels below are anonymized and are not in the order listed earlier in this card; they illustrate the spread across classes, not the identity of any class):

Class	LODO F1	LOSO F1
Class A	0.97	0.95
Class B	0.87	0.71
Class C	0.81	0.76
Class D	0.83	0.71
Class E	0.69	0.50

Summary

The background/normal-air condition is recognized robustly across new devices and new sites. Distinct-source hazards travel reasonably well to unseen conditions. One class remains materially weaker under the strictest cross-site test. The gap between the in-domain (approx 0.88) and leave-one-site-out (~0.73) numbers is the honest cost of moving to a new environment, and is why per-installation calibration is part of the recommended deployment (not provided here).

Citation & Lineage

This model is developed on the SNT CNT-chemiresistor sensing platform. The seminal work on gas sensitivity of chemiresistors based on modified CNT is described in:

Panes-Ruiz, L. A., Shaygan, M., Fu, Y., Liu, Y., Khavrus, V., Oswald, S., Gemming, T., Baraban, L., Bezugly, V., and Cuniberti, G. (2018). Toward Highly Sensitive and Energy Efficient Ammonia Gas Detection with Modified Single-Walled Carbon Nanotubes at Room Temperature. ACS Sensors, 3, 79. https://doi.org/10.1021/acssensors.7b00358

The foundational platform work (industrial-gas identification on functionalized nanomaterial chemiresistors) is described in:

Huang, S., Croy, A., Panes-Ruiz, L. A., Khavrus, V., Bezugly, V., Ibarlucea, B., & Cuniberti, G. (2022). Machine Learning-Enabled Smart Gas Sensing Platform for Identification of Industrial Gases. Advanced Intelligent Systems, 4(4), 2200016. https://doi.org/10.1002/aisy.202200016

@article{Huang2022SmartGasSensing,
  title   = {Machine Learning-Enabled Smart Gas Sensing Platform for Identification of Industrial Gases},
  author  = {Huang, Shirong and Croy, Alexander and Panes-Ruiz, Luis Antonio and Khavrus, Vyacheslav and Bezugly, Viktor and Ibarlucea, Bergoi and Cuniberti, Gianaurelio},
  journal = {Advanced Intelligent Systems},
  volume  = {4},
  number  = {4},
  pages   = {2200016},
  year    = {2022},
  doi     = {10.1002/aisy.202200016},
  publisher = {Wiley}
}

Note: the cited paper establishes the sensing platform and its provenance; it does not describe this specific fire/smoke classifier.

Model Card Authors & Contact

• Authors: SmartNanotubes Technologies (SNT)
• Contact: info@smart-nanotubes.com

How to Get Started with the Model

Model weights are not provided. This card documents the model's capabilities and validated performance.

For evaluation access, integration, or partnership discussions, please contact SmartNanotubes Technologies via info@smart-nanotubes.com.