Fig. 4: Biomimetic olfactory lightweight machine-learning strategy for gas type identification and odor concentration prediction.

a Scheme of the neural network architecture used for gas type identification and odor concentration prediction from olfactory information in a simulated scenario. b Confusion matrix of the gas type identification strategy when recognizing eight environmental gases (including atmosphere, ethanol, acetone, methanol, ammonia, carbon oxide, hydrogen sulfide, and methanal). c Concentrations of the odor predicted by a biomimetic neural network algorithm. The data number refers to the number of data points used for algorithm processing. d Overall accuracy of the odor concentration prediction methods for comparison. The proposed network shows the best accuracy among the tested strategies. n = 10 for each group. The error bars denote standard deviations of the mean. e Prediction accuracy versus the number of sensors. The accuracy increases with more sensors. The illustration is a scheme showing damage to random parts of a gas sensor array. f Representative set of examples during the recognition of gas type and prediction of odor concentration in a practical environment, including a standard environment and humidity interference. The odor labels and corresponding classification vectors from the olfactory inputs are shown