Research on Gas Recognition Based on Stacked Denoising Autoencoders

The gas data coming from an array of chemical gas sensors is a kind of multivariate time-series. This data set is extremely difficult and complex to interpret for human experts. It needs designing hand-made features when applying traditional shallow machine learning algorithms in gas recognition. A new gas recognition method based on Deep Learning were proposed in this paper. It is one of unsupervised feature learning methods that can extract self-adapting features from the gas data, overcoming the complex process in designing features by hands and making the features more general. In this work, two methods based on UCI Machine learning database respectively were compared in the experiments. One of them is a two-hidden-layer structure of deep neural network-Stacked denoising Autoencoders and another is a kind of shallow machine learning algorithms. The results show that extracting features automaticly using Deep Learning is a simpler and more universal way in gas recognition. The method proposed in this paper not only improves the gas classification accuracy, but also reduces complexity of the process in shallow machine learning alogithms, so it is valuable to be applied in practice.