Towards Robust Odour Recognition Systems: Algorithm and Hardware Design for Multimodal Odour-Image Classification and Continual Learning

Odour recognition using electronic noses (EN) focuses on interpreting the response signal of gas sensors in response to an odour stimulus. However, this task is challenging due to the non-specificity of the low-cost resistive sensors and the problem of sensor drift. These issues are especially pronounced when transitioning from controlled laboratory environment to real-world, open-air scenarios, where the constantly changing environment affects the recognition accuracy of the algorithm. This thesis investigates three strategies to enhance odour recognition: multimodal sensing, continual learning and hardware design.
This thesis proposes that some of these issues can be mitigated if there is extra information available. Combining odour and image modality arises as a method of including more information in odour recognition. To investigate this, a dataset on fruit ripening was collected, featuring several case studies of increasing complexity. This work explores how the two modalities can complement each other, and it demonstrates that the accuracy of odour classification can be improved with odour-image decision fusion strategies.
Another approach for improving EN applications is transforming the scenario from static to continual learning. Most odour recognition methods are offline, meaning that the odour recognition algorithm is trained once and deployed after. In this work, a continuous test-time adaptation approach is explored to demonstrate that the accuracy of the classification can be improved both with unlabelled and active labelled data when following a self-training paradigm.
In practice, both approaches need to run on an embedded system, as a typical EN application requires it to be portable. This work presents hardware designs for edge FPGA-CPU adaptive System-on-Chip platforms, implemented with high-level synthesis, for running the training of fully connected networks and 1D convolutional networks with image inference. The resulting systems achieve low-latency performance suitable for portable EN applications and other applications in the time series classification domain.