Towards Gas Discrimination and Mapping in Emergency Response Scenarios Using a Mobile Robot with an Electronic Nose

Han Fan, Victor Manuel Hernandez Bennetts, Erik Schaffernicht, Achim J. Lilienthal

TL;DR

This paper presents an innovative mobile robot system equipped with an "electronic nose" for real-time gas discrimination and spatial mapping in emergency response scenarios. By employing unsupervised, online learning, the system can identify and map unknown chemical compounds, providing crucial situational awareness for first responders in hazardous environments.

What Problem Does This Paper Address?

Emergency responders, such as firefighters and hazmat teams, often face environments contaminated with unknown hazardous airborne chemicals. Traditional gas sampling methods are too slow and complex for these dynamic situations. Existing robotic solutions struggle with identifying unknown substances and mapping their distribution in real-time, especially when prior information about the chemicals present is unavailable. This paper tackles the challenge of developing a robot that can autonomously detect, discriminate, and map multiple chemical compounds in real-time, even when their identities are unknown, to aid in emergency response.

Methods

The core of the system is a three-stage pipeline:

  • Gas Detection: An ensemble of one-class classifiers (OCGM and OCNN) is used to detect the presence of gases and compensate for sensor drift. This module learns from baseline responses (clean air) and then identifies deviations, triggering further analysis.
  • Gas Discrimination: This module employs an unsupervised, online learning approach based on a modified KmP algorithm. It automatically determines the number of chemical compounds present and learns a probabilistic classifier to distinguish them. Crucially, this learning is "event-triggered," meaning it only occurs when gas is detected, conserving computational resources. The system also innovatively determines the search space for the number of chemical classes automatically, removing a manual parameter.
  • Gas Distribution Mapping: The class posteriors from the discrimination module are fed into a Multi-Compound (MC) Kernel DM+V algorithm. This generates spatial classification maps that visualize the likelihood of different chemical compounds being present at various locations within the environment. This mapping is performed online during the robot's exploration, providing immediate situational awareness.

Datasets

The experiments were conducted in a simulated emergency environment (firefighter training facility, laboratory rooms, and a basement). The "datasets" were generated by releasing commercially available liquids: ethanol (95% pure), 1-propanol (99.5% pure), and acetone (100% pure) from beakers. The UWAR nose, a prototype array of MOX sensors, was used to collect the sensor responses. The paper does not introduce a new benchmark dataset but rather uses real-world experimental data generated by these controlled releases.

Key Insights and Ideas

  • Unsupervised Online Learning is Crucial: The ability to learn and adapt in real-time without prior knowledge of chemical identities is paramount for emergency response. Supervised methods are insufficient when the "what" is unknown.
  • Event-Triggered Learning for Efficiency: By only updating models when gas is detected, the system significantly reduces computational load, making it feasible for resource-constrained mobile robots.
  • Automatic Search Space Determination: Eliminating the need for users to specify the range of possible chemical compounds (the search space for K*) makes the system more robust and user-friendly in unknown scenarios.
  • Mapping Class Posteriors, Not Just Concentrations: Mapping the probability of a substance being present is more reliable than mapping absolute concentrations when the exact chemical identity and its sensor response characteristics are not fully known.
  • Sensor Drift Compensation is Essential: MOX sensors are prone to drift, and the system's ability to detect and compensate for this drift through adaptive baseline modeling is critical for accurate gas detection.

Key Results

  • High Accuracy in Gas Discrimination: The online discrimination model achieved high accuracy (H values generally above 0.90) in classifying measurements during exposure states across multiple experimental trials.
  • Significant Data Reduction: The online KmP approach reduced the amount of data processed for model updates by at least 54.04% on average compared to an offline approach.
  • Accurate Gas Distribution Maps: The generated classification maps accurately indicated the presence of chemical analytes near their respective sources, with areas of high posteriors correlating well with the known locations of ethanol, propanol, and acetone.
  • Successful Identification of Multiple Compounds: The system could discriminate between different chemical compounds and map their distributions, even in more complex scenarios with multiple sources.

Limitations and Future Work

  • Separation of Close Sources: The system's ability to reliably separate sources of different chemicals is limited when they are very close together, potentially leading to misclassification of gas mixtures. This is particularly challenging if a baseline state (clean air) cannot be reached between the sources.
  • Missed Gas Detection Events: In some instances, the gas detection module missed low-concentration gas events, which were later identified by other algorithms.
  • Future Work:
    • Integrate with novel change point detection algorithms (e.g., TREFEX) for improved detection of low-concentration events.
  • Develop one-class classifiers for specific chemical analytes to supplement the unsupervised multi-class approach, which could be beneficial for repetitive detection of the same unknown gas.
  • Extend the framework to other robotic platforms (drones, multi-robot systems) and explore new issues that arise.
  • Investigate methods to estimate the minimum separation distance for reliable source identification and to detect gas mixtures.
  • Develop more standardized testbeds for repeatable in-field experiments and simulations.

Summary updated

BibTeX
@inproceedings{fan2019towards,
  author = {Han Fan and Victor Manuel Hernandez Bennetts and Erik Schaffernicht and Achim J. Lilienthal},
  title = {{Towards Gas Discrimination and Mapping in Emergency Response Scenarios Using a Mobile Robot with an Electronic Nose}},
  booktitle = {Sensors},
  year = {2019},
  url = {https://openreview.net/forum?id=OLBkT2LdHs},
}
Figure