Fast Measurement With Chemical Sensors Based on Sliding Window Sampling and Mixed-Feature Extraction

In many application scenarios of chemical sensors, it is vital to obtain the sensing results rapidly. An economic and effective solution for that is through predicting the final value from the earlier transient response. However, conventional approaches require large amount of measurement data to build relationship between the concentration and the feature of the transient response. In this work, a systematic scheme including data collection and prediction is proposed to minimize both the required time and amount of measurement data for training, while achieving high prediction accuracy. A sliding window sampling approach with data augmentation is developed to generate more sequence sets for training from limited amount of measurement data. With this method, any small segment on the early transient response curve can be used for data sampling and test, enabling ease of implementation for practical applications. Further, a model combining long short-term memory (LSTM) neural network and polynomial fitting is designed for mixed feature extraction with less required data, which are then input to multilayer perceptron for prediction. The scheme is applied to real measurement data, showing significant improvement of prediction accuracy compared to previous methods.