Anti-Drift Gas Detection Algorithm Based on Neural Network

Recently, long-term gas detection has attracted much attention due to its being a key factor for electronic nose (E-Nose) applications. However, the sensor drift effect can significantly reduce the performance of the sensor. Therefore, in this work, we proposed a new drift compensation method by optimizing feature selection, model construction, and training methods to study drift-resistant gas detection based on convolutional neural network (CNN) methods. First, the attention mechanism is used to screen the specific features of the gas data and remove the low-weight features. Moreover, a multiscale feature extraction network is designed so that the features fused by the three-layer convolution are used as the final classification feature input to extract the depth features keeping the drift unchanged. Simultaneously, the segmented training method and the targeted cyclic training model are adopted to reduce the required experimental data. Importantly, based on the largest gas drift dataset currently, the proposed method maintains the average gas detection accuracy beyond 80% in three years, and the long-term stability of gas detection is effectively improved. Therefore, our findings provide an effective way to solve the sensor drift effect.