A machine learning field calibration method for improving the performance of low-cost particle sensors
A machine learning field calibration method for improving the performance of low-cost particle sensors
复制标题
DOI：
10.1016/j.buildenv.2020.107457
复制
发表时间：
2021-01-22
影响因子：
7.4
通讯作者：
Boor, Brandon E.
中科院分区：
文献类型：
--
作者：
Patra, Satya S.;Ramsisaria, Rishabh;Boor, Brandon E.
使用低成本光学粒子计数器 (OPC) 测量建筑物中的空气颗粒物通常不准确且存在不确定性。本研究介绍了一种通过机器学习提高低成本 OPC 测量室内颗粒物性能的方法。在一座有人居住的净零能源房屋中进行了为期两个月的现场测量活动。所研究的 OPC（OPC-N2，Alphasense Ltd.）报告尺寸分级浓度为 0.38 至 17.5 μm。位于同一位置的参考仪器包括扫描迁移率粒度仪（SMPS：0.01-0.30μm）和光学粒度仪（OPS：0.30-10μm）。机器学习现场校准方法应用高斯过程回归 (GPR)，包括两个部分：(1.) 将尺寸分辨 OPC 计数效率从 0.38 校正到 10 μm，以及 (2.) 预测低于 OPC 0.38 μm 检测限的体积尺寸分布（质量代理）。现场校准方法适用于报告大小分级浓度的 OPC。在(1.)中，使用 GPR 功能以 OPS 作为参考来校正 0.38 至 10 μm 之间 OPC 的尺寸分辨计数效率。在 (2.) 中，使用 SMPS/OPS 作为参考，使用第二个 GPR 函数来预测低于 0.38 μm 的体积尺寸分布。考虑到 0.38 μm 以下的颗粒在累积模式下对体积浓度的显着贡献，因此实现了这一点。与 SMPS/OPS 相比，机器学习现场校准方法显着提高了 OPC 报告的尺寸积分体积浓度（PV2.5、PV10）的准确性。皮尔逊系数有所改善（校正前：0.59-0.83；校正后：0.98-0.99）；决定系数（校正前：0.35-0.69；校正后：0.97-0.98）；和平均绝对百分比误差（校正前：35-69%；校正后：19-25%）。
Measurements of airborne particles in buildings with low-cost optical particle counters (OPCs) are often inaccurate and subject to uncertainties. This study introduces a methodology to improve the performance of low-cost OPCs in measuring indoor particles through machine learning. A two-month field measurement campaign was conducted in an occupied net-zero energy house. The studied OPCs (OPC-N2, Alphasense Ltd.) report size fractionated concentrations from 0.38 to 17.5 mu m. Co-located reference instrumentation included a scanning mobility particle sizer (SMPS: 0.01-0.30 mu m) and an optical particle sizer (OPS: 0.30-10 mu m). The machine learning field calibration method applies Gaussian Process Regression (GPR) and includes two components: (1.) correction of the size-resolved OPC counting efficiency from 0.38 to 10 mu m and (2.) prediction of volume size distributions (mass proxy) below the 0.38 mu m detection limit of the OPC. The field calibration method is applicable to OPCs that report size fractionated concentrations. In (1.), a GPR function was used to correct the size-resolved counting efficiency of the OPCs between 0.38 and 10 mu m using the OPS as reference. In (2.), a second GPR function was used to predict the volume size distribution below 0.38 mu m using the SMPS/OPS as reference. This was done given the significant contribution of sub-0.38 mu m particles to volume concentrations in the accumulation mode. The machine learning field calibration method resulted in a significant improvement in the accuracy of size-integrated volume concentrations (PV2.5, PV10) reported by the OPCs as compared to the SMPS/OPS. Improvements were seen in the Pearson coefficient (before correction: 0.59-0.83; after correction: 0.98-0.99); coefficient of determination (before correction: 0.35-0.69; after correction: 0.97-0.98); and mean absolute percentage error (before correction: 35-69%; after correction: 19-25%).