Adaptive chemical agent detection in dynamic, changing environments
Anna Ladi · White Rose eTheses Online (University of Leeds, The University of Sheffield, University of York) · 2015
In this thesis a framework for adaptive chemical detection is developed, considering the application scenario of autonomous, robot mounted chemical agent detection in dynamic, changing environments. Chemical detection is performed by the Receptor Density Algorithm (RDA), a previously developed immune-inspired anomaly detection algorithm, which suffers from a decrease in its performance when the background environment changes. Focusing on the software part of the system, the goal of this thesis is to adapt the RDA quickly and autonomously, without requiring user feedback. The approach adopted is to first detect a change in the background data generating distribution, also defined as concept drift, and adapt in response to this detected change. Statistical hypothesis testing is used to determine whether there has been concept drift in consecutive time windows of the incoming sensor data. Five different statistical methods are tested on mass spectrometry data, enhanced with artificial concept drift that signifies a changing environment. The results show that, while no one method is universally best, statistical hypothesis testing can detect concept drift in the context of chemical sensing and it can differentiate between anomalies and concept drift. The adaptation of the system, which is triggered by the detection of concept drift, is achieved by switching to an ensemble (a set) of RDAs , created from a pool of pre-existing candidates. A novel mechanism for evaluating and selecting the members of the ensemble from this pool is proposed; it uses implicit performance information, extracted from the dynamics of the RDA, and does not require new user input to evaluate the candidates for the new environment. An ensemble of 5 members, selected in this way is found to be significantly better than a single RDA, the previously known best, reducing both the false detections and the number of missed anomalies. This method for selecting the ensemble members is also found significantly better than populating the ensemble based on their performance of the environment before concept drift. Finally, it is found that the ensemble can be created online, with its performance converging to the offline variant.