Neural Network for Odor Sensor Array Spike Encoding Inspired by Mammalian Olfaction

Signal processing in electronic nose (e-nose) odor detection involves a series of complex and time-consuming steps, such as noise reduction, filtering, normalization in signal preprocessing, as well as feature generation, selection, and dimensionality reduction. When dealing with different gas sensors, these processing methods might require meticulous adjustments, and each sensor might need own specific processing strategy. Such strategies are often tedious and overly reliant on intricate feature engineering, leading to time-consuming operations and poor reproducibility. To address this issue, this paper proposes a biomimetic olfactory perception model inspired by the mammalian olfactory system. Unlike traditional e-nose methods, our model deeply simulates the core structures and functions of the mammalian olfactory system, especially the biomimetic design for neural conduction and information encoding. This allows the model to directly handle the raw sensor responses of the e-nose, eliminating the conventional preprocessing and feature selection steps. The experimental results show that our biomimetic model achieved an average recognition rate of 95.8% on a traditional Chinese medicine dataset, significantly outperforming the traditional method's 90.84%. The model not only demonstrates excellent feature extraction capabilities but also shows extreme robustness when the types and numbers of sensors change, enhancing the reliability and accuracy of the electronic nose in practical applications.