A mixed-kernel, variable-dimension memristive CNN for electronic nose recognition

Jiahao Chen, Lidan Wang, Shukai Duan

Abstract: Due to the dynamic characteristics, memristors have great potential for implementing various neural network training and applications. By applying memristors to neural networks as a strategy, the hardware architecture can be taken full advantage of to accelerate matrix operations. A mixed-kernel, variable-dimension memristive convolutional neural network (MixVMCNN), which can identify and classify six gases detected by the electronic nose, is proposed. To generate better gas recognition technique, we provide various receptive fields for the network by using different sizes of convolutional kernels in the same layer. We propose a novel variable-dimension convolutional approach that extracts features in different dimensions. Furthermore, we make use of memristor arrays to make the most of the hardware structure to speed up calculation. Experimental results indicate that the designed network, using only 56 weights, achieves the electronic nose gas classification with high accuracy of 99.72%, based on the Chemical gas sensor array dataset. To further exam the effectiveness of the model, supplemental experiments show that this model has stronger noise-robustness than other networks.