Research On Calibration Approaches Of Air Monitoring Sensors

In recent years,air pollution has grown into a global issue.Air pollution monitoring is an important means to deal with the problem.The government has established many standard air monitoring stations to monitor air pollution in real time.The data accuracy is high,but the construction costs of the stations are also high.The deployment is relatively sparse.Low-cost micro sensor devices are often used to build large sensor networks to achieve intensive regional monitoring.However,due to factors such as cross-interference and sensor aging,there is certain deviation between the reading of the micro sensor device and the standard concentration.In order to ensure the data quality of the sensors in the network,these sensor devices need to be calibrated,and the calibration task faces great challenges.To address the above challenges,this paper proposes a category-based calibration model and a time series-based calibration model.The contributions of the paper are summarized as follows.?1?This paper proposes a category-based calibration approach?CCA?.Traditional calibration models usually use only one regression model to construct the mapping relationship between sensor readings and standard concentration.CCA builds multiple regression models based on the concentration category of pollutants to build a more accurate mapping from sensor readings to reference values.In addition,CCA introduces two fault tolerance modules:classification tolerance and sample tolerance.The former alleviates the effects of misclassification of classification models,and the latter improves the robustness of each regression model.?2?This paper proposes a time series-based calibration model?C-Net?.The traditional calibration model is a point-to-point structure,it is often difficult to mine the dependencies in time series data.C-Net is an attention-based encoder-decoder structure.The attention mechanism is in the encoder stage,and it will select the relevant feature sequence at each time step according to the previous hidden state of the encoder to learn the cross-interference feature.In addition,C-Net also has a trend feature extraction module to extract the change trend of target pollutants.The encoder's RNN model encodes the extracted features into fixed-length time-series encoding vectors.The decoder receives the encoded vector and the historical reference values to obtain the final calibration value.C-Net can mine effective timing characteristics and cross-interference characteristics from limited input data,thereby improving the calibration accuracy.The above models are evaluated on the real data sets of carbon monoxide?CO?and ozone?O₃?in two cities?Lanzhou and Fuzhou?.Experiments show that CCA can effectively improve the performance of the regression model,and its fault-tolerant strategy guarantees the robustness of CCA.C-Net outperforms traditional regression-based methods on most data sets.