Study On Clustering-Based Sensory Data Calibration For Air Pollution Sensing System

To deal with severe air pollution,the sensor-based air pollution sensing system has gradually become a complement for existing standard air pollution monitoring architecture.Because the sensing system can realize large-scale deployment to obtain air pollution information of high spatial-temporal resolution at a relatively low cost.It is significant for fine-grained monitoring and treatment on air pollution.However,due to the performance limitation of low-cost sensors,disadvantages such as cross sensitivity and sensor drift,must result in deviations between sensory data and the ground truth.Meanwhile,there are many sensing nodes in the sensing system,but the reference data that can be used for calibration are relatively scarce.It is difficult for traditional calibration methods to perform calibration operations efficiently.The premise of effective treatment is reliable sensory data of air pollution.To guarantee the data quality of the sensing system,an accurate and efficient sensing system calibration scheme is essential.Technically with insufficient prior knowledge,determining calibration model structure is challenging using a global fitting model because the mechanism of interference factors is complex.Although ANN-based methods ignore the complex conditions above,they also have disadvantages regarding generalization,interpretability and calculation cost.This paper proposes a multi-segment regression calibration model（CSR）for a single sensor node,reducing the measurement deviation from temperature and relative humidity of the particulate matter sensor.Based on this,a system-level calibration model KNN-CSR for the particulate matter sensing is designed,and the performance of the model is validated by practical data experiments.The main work is as follows:（1）A cluster-based multi-segment regression calibration model（CSR）is proposed.The CSR model uses clustering algorithms to divide samples into several clusters adaptively and each cluster has an individual multiple linear regression model.Multi-segment regression is intended to approximate an ideal global fitting with few cost.The CSR model requires a few training samples and does not rely on much prior knowledge.It is demonstrated by experiments that the calibration results of the CSR model has been greatly improved.For two target pollutants,PM_2.5 and PM₁₀,compared with the traditional global multiple linear regression model,the error is reduced by at least 14.97%and 9.68%with higher stability.（2）A calibration model KNN-CSR of the sensing system is designed on the basis of the CSR model.With scarce reference data,the model from CSR is transferred to other sensing nodes in the sensing system for calibration.The KNN algorithm is applied to classify the data from sensing nodes to be calibrated in the sensing system and to match the corresponding calibration model for calculation.It is validated that the KNN-CSR calibration model can greatly reduce the data error of sensing nodes data by transferring,and its performance is close to the in-field calibration with multiple linear regression under sufficient reference data.It has great significance for air pollution sensing system in long-term deployment.The data used in this study comes from a practical air pollution sensing system.The calibration scheme has realized software deployment on the sensing terminal and data platform of the system,continuously maintaining and improving sensory data quality.