Achim J. Lilienthal
Gas Distribution Mapping and Gas Source Localisation with a Mobile Robot
PhD Thesis, University of Tübingen, December 2004
Abstract: This thesis addresses two fundamental problems concerning the application of electrochemical gas sensors on a mobile robot in a real world environment: gas distribution mapping and localisation of a static gas source.
Apart from the limitations of current sensor technology, the main difficulty for both of these tasks results from the spreading of gases under natural conditions. Diffusion plays only a minor role in the distribution of odourant molecules at room temperature. The dominant transport mechanisms are convection flow and turbulence. As a consequence, the resulting concentration field is patchy and chaotic, and the actual gas source is usually not located at the point of highest concentration.
In order to create a map of a gas distribution it is thus required to extract time-constant properties of the concentration field from a series of sensor readings collected by a mobile robot. In contrast to conventional mapping techniques, where a considerable overlap between single measurements can be assumed, gas sensor measurements provide information about a very small area. The problem of gas distribution mapping is therefore to create a representation of the average concentration field from sparse point samples of the instantaneous distribution with little or no overlap between single measurements. A new algorithm is introduced in this work to create a representation that stores belief about the average relative concentration of a detected gas in a grid structure.
The problem of gas source localisation can be broken down into three subtasks: gas finding, source tracing and source declaration. Gas finding — the detection of an increased concentration — amounts to a basic search task and the selection of a suitable threshold value, a problem that is not addressed in this work. The main part of this thesis is rather concerned with the latter two aspects: following the cues determined from the sensed gas distribution towards its origin (source tracing) and establishing that the source has been found (source declaration). In order to avoid limitation to an environment with a strong airflow, the investigated strategies do not rely on information about wind direction and speed. Accordingly, the experiments were carried out in an unventilated indoor environment where previously suggested methods for gas source localisation, which include periods of upwind movement, are not applicable due to the limitations of current anemometers.
A straightforward solution for gas source tracing is to follow an instantaneously measured spatial concentration gradient. Such a strategy was realised on a mobile robot by implementing a direct sensor motor coupling in the manner of a Braitenberg vehicle. The particular contribution of this work is a detailed statistical evaluation of the tracing performance based on a large number of experiments. While gradient following is often mislead by transient concentration maxima, it could be shown with high statistical significance that the path length required to reach the source can be reduced on average compared to random search.
As a further gas source tracing method, a biomimetic strategy was investigated that is based on the key elements of the behaviour of male silkworm moths to find a mate guided by sexual pheromones. A modification for use on mobile robots is proposed that does not depend on information about the local wind speed. The performance of this tracing strategy was tested in a largely uncontrolled indoor environment and evaluated by statistical means. Despite the considerably larger size of the robot compared to the moth, the results suggest that the modified strategy decreases the average robot-to-source distance compared to random exploration, because it can keep the robot in the vicinity of a gas source after single gas patches have been discovered by initial exploration.
In order to address the full gas source localisation problem, a tracing strategy has to be extended by an additional declaration mechanism to determine that the source has been found with high certainty. It is generally not sufficient to search for maxima of the instantaneous concentration distribution in order to solve the declaration task. However, it was demonstrated that peaks in sensor response can provide a rough estimate of the gas source location if the search space is restricted. In experiments in a one dimensional scenario, the sensing strategy was found to have a dominant influence. A strong correlation between sensor response and proximity to a source could be obtained only if the robot was driven with a constant, sufficiently high speed, while such a correlation could not be observed if a stop-sense-go strategy was applied. A possible explanation for this effect suggests an alternative feature that might be used to recognise a gas source. Rather than looking for a global maximum of concentration, a gas source can be distinguished by an increased frequency of local maxima. This was also found in experiments in a two dimensional scenario where the sensor-motor connections of the implemented gas-sensitive Braitenberg vehicle were crossed. In this way, the robot performs exploration and concentration peak avoidance, resulting in a path that reflects the frequency of local maxima in the inspected area. A visualisation of this path offers therefore a possible method for gas source declaration without using additional sensors.
Finally, the thesis investigates the possibility to classify a suspected object as being a gas source or not from a pattern in a series of spatially and temporally sampled concentration data. Such a pattern was determined by applying machine learning techniques. The results of this ongoing work demonstrate the feasibility of the approach and show that high classification rates can be achieved using support vector machines. An analysis of the most important features for classification, the dependency of the classification rate on the desired declaration accuracy, and a comparison with the classification rate that can be achieved by selecting an optimal threshold value regarding the mean sensor signal is also presented.
Download:
Thesis (Colour Version): [PDF (33.5MB)]
Thesis (Web Optimized Version): [PDF (15.1MB)]
@PHDTHESIS{Lilienthal:PhdThesis:2004,
  AUTHOR = {Lilienthal, Achim~J.},
  TITLE = {Gas Distribution Mapping and Gas Source Localisation with a Mobile Robot},
  SCHOOL = {Wilhelm-Schickard Institute, University of T{\"u}bingen},
  YEAR = {2004},
  MONTH = {December}
}