Low-dimensional metal oxides-based electronic noses have been applied in various fields, such as food quality, environmental assessment, coal mine risk prediction, and disease diagnosis. However, the applications of these electronic noses are limited for conditions such as precise safety monitoring because electronic nose systems have problems such as poor recognition ability of mixed gas signals and sensor drift caused by environmental factors. Advanced algorithms, including classical gas recognition algorithms and neural network-based algorithms, can be good solutions for the key problems. Classical gas recognition methods, such as support vector machines, have been widely applied in electronic nose systems in the past. These methods can provide satisfactory results if the features are selected properly and the types of mixed gas are under five. In many situations, this can be challenging due to the drift of sensor signals. In recent years, neural networks have undergone revolutionary changes in the field of electronic noses, especially convolutional neural networks and recurrent neural networks. This paper reviews the principles and performances of typical gas recognition methods of the electronic nose up to now and compares and analyzes the classical gas recognition methods and the neural network-based gas recognition methods. This work can provide guidance for research in related fields.

In recent years, the application of Deep Neural Networks to gas recognition has been developing. The classification performance of the Deep Neural Network depends on the efficient representation of the input data samples. Therefore, a variety of filtering methods are firstly adopted to smooth filter the gas sensing response data, which can remove redundant information and greatly improve the performance of the classifier. Additionally, the optimization experiment of the Savitzky–Golay filtering algorithm is carried out. After that, we used the Gramian Angular Summation Field (GASF) method to encode the gas sensing response data into two-dimensional sensing images. In addition, data augmentation technology is used to reduce the impact of small sample numbers on the classifier and improve the robustness and generalization ability of the model. Then, combined with fine-tuning of the GoogLeNet neural network, which owns the ability to automatically learn the characteristics of deep samples, the classification of four gases has finally been realized: methane, ethanol, ethylene, and carbon monoxide. Through setting a variety of different comparison experiments, it is known that the Savitzky–Golay smooth filtering pretreatment method effectively improves the recognition accuracy of the classifier, and the gas recognition network adopted is superior to the fine-tuned ResNet50, Alex-Net, and ResNet34 networks in both accuracy and sample processing times. Finally, the highest recognition accuracy of the classification results of our proposed route is 99.9%, which is better than other similar work.