A portable ultrafine particulate matter sensor calibration system

Abstract

The present application relates to a kind of portable ultrafine particle sensor calibration system, including standardization pneumatic intake and shunt module, multi-dimensional physical and environmental perception module, data signal pre-processing module and double-layer machine learning calibration module, standardization pneumatic intake and shunt module includes cutting head, silica gel drying tube and multi-way shunt, multi-dimensional physical and environmental perception module includes temperature humidity sensor, SMPS and portable sensor, data signal pre-processing module includes original data receiving module, diffusion loss compensation module and feature matrix construction module, original data receiving module receives the data of original data acquisition module and environmental parameter acquisition module, the present system solves the systematic overestimation problem that microsensor direct-reading lung deposition surface area macro health exposure index is prone to appear, for carrying out high-resolution personal mobile exposure assessment and urban grid environmental decision-making provides highly reliable bottom data support.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring and aerosol sensing data processing technology, specifically relating to a portable ultrafine particulate matter sensor calibration system. Background Technology

[0002] With the rapid expansion of megacities worldwide, ultrafine particulate matter (UFPs) in ambient air (particles with an aerodynamic diameter of less than 100 nm) has become a key research focus in epidemiology and public health due to its large specific surface area and ability to penetrate the human respiratory barrier. In recent years, lung deposition surface area (LDSA) has been considered a more representative indicator of nanoscale exposure than traditional mass concentration due to its high biological relevance to particulate matter-induced toxicological responses. However, traditional benchmark instruments for high-precision particulate number-size distribution (PNSD) measurements (such as scanning electromobility particle size spectrometers, SMPS) are bulky, expensive, and require strict temperature and humidity control, making them unsuitable for high-density gridded monitoring. Against this backdrop, miniature portable sensors based on the unipolar diffused charging principle have been widely applied in urban micro-monitoring networks. While these miniature sensors offer significant advantages in engineering deployment, their measurement reliability often faces severe physical and algorithmic challenges when conducting long-term monitoring in the complex and variable urban atmospheric environment.

[0003] Existing portable sensors suffer from structural defects at both the physical sampling and internal algorithm inversion stages that are difficult to overcome through a single approach. During the physical air intake and transport stages, drastic temperature and humidity changes in the ambient air mass can distort the aerosol state entering the measurement chamber. In particular, the hygroscopic growth of nanoparticles caused by high humidity directly interferes with the charging efficiency of unipolar ions. Simultaneously, highly diffusive ultrafine particles are prone to wall collision losses in poorly regulated pipelines. Inside the sensor's measurement chamber, particles of different sizes exhibit strong spatial hydrodynamic heterogeneity. For larger particles (e.g., 300 nm), they are easily pushed towards the pipeline edge by radial electrostatic forces upon entering the downstream, resulting in aerodynamic stagnation and cutoff losses in the chamber's vortex region. The sensor's built-in concentration inversion matrix, based on the assumption of uniform spatial mixing and ideal electromigration penetration, cannot identify the actual physical losses caused by flow field distortion. To balance weak characteristic current signals, the internal algorithm triggers mathematical misjudgment and overcompensation during matrix solving, passively assigning excessively high computational weights to large-particle channels, leading to severely overestimated measurements. On the other hand, for nucleated particles with extremely small diameters (such as 10 nm), under clean operating conditions with low total aerosol concentrations, the effective microcurrent signal generated by the particles is close to the detection limit of the hardware electrometer. Due to the extremely low signal-to-noise ratio, the unavoidable electronic thermal noise and circuit zero-point drift within the instrument are easily misjudged as effective signals by the inversion algorithm, thereby triggering a noise amplification effect and reversing artificially high abnormal peak values ​​of number concentration at the extremely small particle size end.

[0004] The aforementioned chain of errors, from the underlying "physical flow field loss" to the upper-level "algorithm misjudgment amplification," not only leads to jagged distortions in the local particle size distribution but also causes a nonlinear, systematic overestimation of the geometric mean particle size (GMD) and the core health exposure indicator, lung deposition surface area, on a macroscopic level. Directly adopting such uncorrected, unfiltered data for gridded exposure assessments would severely exaggerate the health risks of acute exposure during periods of high pollution, while masking the actual improvement in air quality under clean conditions, thus completely disrupting the accuracy of epidemiological studies. Current sensor error calibration techniques largely rely on simple multiple linear regression (MLR) or static empirical formulas, making it difficult to isolate the high-dimensional nonlinear interference of complex meteorological parameters. Some models employing general machine learning treat calibration as a purely "black box" data fitting process, ignoring the continuity constraint of the natural log-normal distribution of aerosols and failing to isolate and specifically correct the "flow field stagnation" of large particles and the "noise amplification" of small particles within the sensor, resulting in poor model generalization ability. Therefore, there is an urgent need for a sensor calibration system that can integrate front-end physical boundary specifications and back-end underlying physical mechanism constraints to achieve high-precision data reconstruction across scales. Summary of the Invention

[0005] The purpose of this invention is to overcome the measurement deviations of existing micro and ultrafine particulate matter sensors in complex atmospheric environments caused by front-end aerodynamic flow field interception loss, environmental interference, and back-end inversion algorithm mathematical overcompensation. This invention provides a portable ultrafine particulate matter sensor calibration system that achieves high-precision digital reconstruction from front-end physical sampling boundary specifications to back-end underlying physical mechanism constraints.

[0006] To achieve the above objectives, the specific technical solution adopted by the present invention is as follows:

[0007] A portable ultrafine particulate matter sensor calibration system includes a standardized pneumatic air intake and splitting module, a multidimensional physical and environmental sensing module, a data signal preprocessing module, and a two-layer machine learning calibration module. The output of the standardized pneumatic air intake and splitting module is connected to the input of the multidimensional physical and environmental sensing module, the output of the multidimensional physical and environmental sensing module is connected to the input of the data signal preprocessing module, and the output of the data signal preprocessing module is connected to the input of the two-layer machine learning calibration module.

[0008] The standardized pneumatic air intake and splitting module includes a cutting head, a silicone drying tube, a bypass air pump, and a multi-way splitter. The cutting head has an air inlet on one side, and the other side is connected to the input end of the silicone drying tube and the bypass air pump. The bypass air pump has a flow rate of 3.0 L / min - 15.0 L / min to maintain the efficient and constant cutting flow rate required by the front-end cyclone cutting head. The output end of the silicone drying tube is connected to the input end of the multi-way splitter. Ambient air is connected to the air inlet of the cutting head through a horizontally oriented air passage pipe.

[0009] The multidimensional physical and environmental sensing module includes a temperature and humidity sensor, SMPS, and a portable sensor. The output of the multi-channel splitter is connected in parallel with two independent gas path branches. The first gas path is connected to the air inlet of the portable sensor, and the sampling flow rate is set to 1.0 L / min - 4.0 L / min. The second gas path is connected to the scanning electromobility particle size spectrometer during the system verification and training phase, and the sampling flow rate is set to 1.0 L / min.

[0010] The multidimensional physical and environmental sensing module also includes a raw data acquisition module and an environmental parameter acquisition module. The environmental parameter acquisition module is connected to a temperature and humidity sensor, and the raw data acquisition module is connected to SMPS and a portable sensor.

[0011] The data signal preprocessing module includes a raw data receiving module, a diffusion loss compensation module, and a feature matrix construction module. The raw data receiving module receives data from the raw data acquisition module and the environmental parameter acquisition module. The output of the raw data receiving module is connected to the input of the diffusion loss compensation module, and the output of the diffusion loss compensation module is connected to the input of the feature matrix construction module. The data signal preprocessing module receives raw aligned data from the multidimensional physical and environmental sensing module. First, for the physical pipeline deposition of particles, the diffusion loss compensation module performs physical attenuation correction, and the calculated correction coefficient is used to compensate for the initial readings of each characteristic particle size channel. The diffusion loss compensation module extracts the characteristic concentration of each particle size channel of the portable sensor, and the ambient temperature, relative humidity, and time synchronously acquired by the temperature and humidity sensor are matrix-stitched by the feature matrix construction module to complete the preparation of the high-dimensional input feature matrix.

[0012] The two-layer machine learning calibration module includes a multi-output random forest backbone module, with the output of the feature matrix construction module connected to the input of the multi-output random forest backbone module; the error output generated by the two-layer machine learning calibration module is connected to the input of the output calibration module; the SMPS connection module is connected to the input of the module training and benchmark comparison module, and the output of the module training and benchmark comparison module is connected to the output calibration module; for core aerosol modes with particle sizes ranging from 10nm to 300nm, the system utilizes a multi-output random forest to simultaneously predict multiple particle size channels to maintain the statistical correlation and spectral continuity between different particle size channels; simultaneously, the system... The system performs local reconstruction of the patch subnetwork for the extreme channels at both ends. Specifically, for the 10nm nucleation channel, it calls the independent random forest submodel with the maximum tree depth to suppress the noise amplification effect under low load conditions. For the 300nm large particle channel, it calls the gradient boost regression submodel with heavy regularization to filter out abnormal responses that may be caused by uneven flow field at the cavity edge, electrostatic migration deviation, or local stagnation. Finally, the system performs corresponding position coverage fusion of the normal channel concentration matrix predicted by the backbone model and the extreme channel concentration matrix predicted by the patch subnetwork to generate and output high-precision particle number and size distribution calibration data.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] (1) Non-black-box calibration algorithm with deep integration of physical mechanisms: It abandons the defects of traditional machine learning that lead to physical distortion due to blind fitting. The system uses a multi-output backbone model to maintain the overall continuity of the core modes and avoids the jagged distortion caused by independent modeling of a single channel.

[0015] (2) Precisely remove instrument-specific physical defects: For the background noise amplification caused by the inferior ion competition and extremely low signal-to-noise ratio of extremely small particles, and the overcompensation caused by the "wall-attached retention of the cavity edge vortex zone" of larger particles, the system specifically uses an independent deep random forest and gradient boosting regression patch subnetwork to repair the extreme false peaks that often occur in complex urban atmospheres without destroying the global distribution.

[0016] (3) Significantly improve the reliability of epidemiological exposure assessment: By dynamically and implicitly compensating for high-dimensional nonlinear meteorological interference, this system solves the problem of systematic overestimation that is easy to occur when micro-sensors directly read macroscopic health exposure indicators such as lung deposition surface area, and provides highly reliable underlying data support for carrying out high-resolution personal mobile exposure assessment and urban grid-based environmental decision-making. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a standardized pneumatic intake and multi-dimensional physical environment sensing hardware structure.

[0018] Figure 2 This is a schematic diagram of the software module structure, which includes a multi-dimensional physical and environmental perception module, a data signal preprocessing module, and a two-layer machine learning calibration module.

[0019] Figure 3 This is a schematic diagram showing the effect comparison of the system of the present invention before and after performing full-process calibration of particulate matter number and particle size distribution.

[0020] Figure 4 This is a schematic diagram illustrating the change in total particulate matter number concentration over time during a typical pollution event using the system of the present invention. Detailed Implementation

[0021] Example:

[0022] To make the objectives, technical solutions, and technical effects of this invention clearer, the following provides several embodiments in conjunction with specific parameter ranges and application scenarios to further illustrate this invention in detail.

[0023] Example 1: System Deployment and Calibration under Typical Urban Winter Conditions

[0024] like Figures 1-2As shown, this invention includes a standardized pneumatic air intake and splitting module and a multi-dimensional physical and environmental sensing module connected sequentially by air paths, as well as a data signal preprocessing module and a two-layer machine learning calibration module for physical constraints that communicate with the multi-dimensional physical and environmental sensing module. The standardized pneumatic air intake and splitting module is structured as follows: an ambient air inlet is connected in series via a horizontally oriented air path pipe to a PM2.5 cyclone cutter head with a rated operating flow rate of 5.0 L / min - 20.0 L / min. One side of the cutter head has an air inlet, and the other side is connected to the input end of a silica gel drying tube and a bypass air pump. The outlet end of the silica gel drying tube is connected to the input end of a multi-path splitter. The output end of the multi-path splitter has two independent air path branches connected in parallel. The first branch is connected to the air intake end of a portable sensor in the multi-dimensional physical and environmental sensing module, with a hardware sampling flow rate set to 1.0 L / min - 4.0 L / min. The second branch is connected to a benchmark scanning electromobility particle size spectrometer during the system verification and training phase, with a sampling flow rate set to 1.0 L / min. The multidimensional physical and environmental sensing module includes a portable sensor, SMPS, a high-precision temperature sensor, and a humidity sensor. The data signal preprocessing module and the two-layer machine learning calibration module for physical constraints constitute the core of the system's computation and control. The data signal preprocessing module receives raw aligned data from the multidimensional physical and environmental sensing module. First, for the physical channel deposition of particles, it uses the diffusion loss compensation module to perform physical attenuation correction. The calculated correction coefficients are used to compensate for the initial readings of each characteristic particle size channel.

[0025]

[0026]

[0027] Where: l(d) p ) represents the diffusion loss rate, d p Indicates particle size.

[0028] After compensation, the module extracts the feature concentrations of the eight particle size channels and performs matrix splicing with the synchronously acquired ambient temperature, relative humidity, and time to complete the preparation of the high-dimensional input feature matrix. The physically constrained two-layer machine learning calibration module receives the high-dimensional input feature matrix and executes an automated calibration forward inference step. Its core operation process is as follows: For the core aerosol mode range with particle size between 10nm and 300nm, the system uses a multi-output random forest to simultaneously predict multiple particle size channels to maintain the statistical correlation and spectral continuity between different particle size channels. At the same time, the system awakens and strengthens the patch sub-network for local reconstruction for the extreme value channels at both ends. Specifically, for the 10nm nucleation state channel, the independent random forest sub-model with the maximum tree depth is called to suppress the noise amplification effect under low load conditions. For the 300nm large particle channel, the gradient boost regression sub-model with heavy regularization is called to filter out abnormal responses that may be caused by uneven flow field at the cavity edge, electrostatic migration deviation, or local stagnation. Finally, the system performs corresponding position coverage fusion of the regular channel concentration matrix predicted by the backbone model and the extreme value channel concentration matrix predicted by the patch sub-network to generate and output high-precision particle number and particle size distribution and lung deposition surface area calibration data that eliminates mathematical overcompensation.

[0029] This embodiment demonstrates a 28-day field measurement of the calibration system under typical urban winter weather conditions. In terms of hardware configuration, the ambient air inlet uses a stainless steel PM2.5 cyclone cutter head with a rated operating flow rate of 5 L / min. The lower end of the cutter head is connected via an 8mm inner diameter TYGON E-3603 flexible hose to a drying tube filled with color-changing silica gel granules and a brushless DC bypass air pump, whose pumping flow rate is locked at 3.0 L / min by a flow controller. The drying tube outlet is connected to a stainless steel three-way multi-channel splitter. The first branch of the splitter connects to a miniature portable ultrafine particulate sensor, whose built-in miniature air pump precisely controls its sampling flow rate at 1.0 L / min; the second branch connects to an SMPS reference instrument, with its sampling flow rate controlled at 1.0 L / min. This combination of hardware parameters ensures that the cyclone cutter head maintains a constant flow rate of 5 L / min. During the full-scale production training phase of the system, the external ambient temperature fluctuated around 5°C, and the relative humidity was around 40%. The preprocessing module first substitutes the diffusion loss physical equation to calculate pipeline loss. For example, for particles of 10.0nm, 16.3nm, and 300.0nm, the diffusion loss ratios are calculated to be 21.35%, 12.01%, and 0.40%, respectively. Corresponding correction coefficients of 1.271, 1.136, and 1.004 are then applied to compensate for the original sensor readings. Subsequently, the algorithm control core initializes a two-layer machine learning architecture, configuring the Global Multi-Output Random Forest (MORF) backbone network parameters as follows: number of decision trees n_estimators = 100, maximum depth max_depth = 10; configuring the 10nm nucleated channel independent random forest model parameters as: maximum depth max_depth = 15; and configuring the 300nm large particle channel gradient boosting regression (GBR) model parameters as: learning rate learning_rate = 0.05, maximum tree depth max_depth = 3. After training and testing, the calibration system with this parameter combination can effectively correct the smoothing effect caused by defects in the built-in mathematical matrix of the sensor, and its calibrated geometric mean particle size (GMD) is highly consistent with the SMPS reference.

[0030] Example 2: Dynamic tracking and reconstruction of extreme pollution cycles under complex atmospheric conditions

[0031] This embodiment uses February 7th to February 10th, 2026 as the calibration condition to verify the system's ability to dynamically track and reconstruct high-frequency spatiotemporal fluctuations. Under this combined condition, the atmosphere is not only accompanied by extreme enrichment of particulate matter during morning and evening traffic rush hours and strong abrupt changes in new particle generation at noon, but also by extreme weather events such as high humidity and even precipitation. Faced with high-humidity external air masses, the silicone drying tube in the standardized pneumatic air intake module suppresses the interference of the high-humidity environment on instrument measurements. At the same time, the multi-dimensional environmental perception module synchronously acquires the characteristics of external high humidity and precipitation and transmits them along with the microcurrent signal to the preprocessing module to generate a feature matrix. During the algorithm calibration execution phase, the two-layer machine learning architecture demonstrates excellent dynamic correction capabilities in the face of complex pollution evolution. When the system experiences a pulse burst of pollutants, a large number of environmental nanoparticles flood in, causing a severe ion competition effect in the measurement chamber; especially when precipitation causes preferential wet deposition of larger particles, the original electrical balance is completely broken, causing extremely small particle sizes such as 10nm to easily induce a noise amplification effect with artificially high values ​​under traditional algorithms. To address this physical phenomenon, the system activated an independent deep random forest patch sub-model (maximum tree depth set to 15) in the nucleation channel. This model, by coupling relative humidity, precipitation identifiers, and time periodic characteristics, successfully suppressed noise amplification under low-load conditions and captured the high-frequency nonlinear mutations in the generation of real new particles. On the other hand, during the pollution accumulation period, a large number of accumulated large particles undergo severe radial electrostatic adhesion and physical cutoff loss in the eddy current region at the edge of the cavity. Faced with weak and distorted characteristic current signals, traditional sensors often exhibit signal passivation or blind mathematical overcompensation. However, the 300nm channel of this system utilizes a heavily regularized gradient boosting regression sub-model (learning rate set to 0.05, maximum tree depth set to 3) to effectively suppress the overestimation tendency caused by incorrect weight allocation based on global historical data, and reasonably restores the 300nm signal that was originally lost due to the adhesion effect. The final calibration output particulate matter concentration curve accurately matches the true dynamic profile of the reference instrument, reproducing the severe spatiotemporal fluctuations during the pollution accumulation period. During this process, the calibration determination coefficients R² for the core exhaust gas particle size ranges such as 42nm and 70nm remained stable between 0.91 and 0.93.

Claims

1. A portable ultrafine particulate matter sensor calibration system, characterized in that... It includes a standardized pneumatic intake and flow splitting module, a multi-dimensional physical and environmental sensing module, a data signal preprocessing module, and a two-layer machine learning calibration module. The output of the standardized pneumatic intake and flow splitting module is connected to the input of the multi-dimensional physical and environmental sensing module, the output of the multi-dimensional physical and environmental sensing module is connected to the input of the data signal preprocessing module, and the output of the data signal preprocessing module is connected to the input of the two-layer machine learning calibration module. The standardized pneumatic air intake and splitting module includes a cutting head, a silicone drying tube, a bypass air pump, and a multi-way splitter. The cutting head has an air inlet on one side, and the other side is connected to the input end of the silicone drying tube and the bypass air pump. The bypass air pump has a flow rate of 3.0 L / min - 15.0 L / min to maintain the efficient and constant cutting flow rate required by the front-end cyclone cutting head. The output end of the silicone drying tube is connected to the input end of the multi-way splitter. Ambient air is connected in series with the air inlet of the cutting head through a horizontally oriented air duct. The multidimensional physical and environmental sensing module includes a temperature and humidity sensor, a SMPS (Supervisory System Power Spectrometer), and a portable sensor. The output of the multi-channel splitter is connected in parallel to two independent gas path branches. The first branch is connected to the inlet of the portable sensor, with a sampling flow rate set to 1.0 L / min - 4.0 L / min. The second branch is connected to a scanning electromobility particle size analyzer during system verification and training, with a sampling flow rate set to 1.0 L / min. The multidimensional physical and environmental sensing module also includes a raw data acquisition module and an environmental parameter acquisition module. The environmental parameter acquisition module is connected to the temperature and humidity sensor, while the raw data acquisition module is connected to both the SMPS and the portable sensor. The data signal preprocessing module includes a raw data receiving module, a diffusion loss compensation module, and a feature matrix construction module. The raw data receiving module receives data from the raw data acquisition module and the environmental parameter acquisition module. The output of the raw data receiving module is connected to the input of the diffusion loss compensation module, and the output of the diffusion loss compensation module is connected to the input of the feature matrix construction module. The data signal preprocessing module receives raw aligned data from the multidimensional physical and environmental sensing module. First, for the physical pipeline deposition of particles, the diffusion loss compensation module performs physical attenuation correction, and the calculated correction coefficient is used to compensate for the initial readings of each characteristic particle size channel. The diffusion loss compensation module extracts the characteristic concentration of each particle size channel of the portable sensor and, together with the ambient temperature, relative humidity, and time synchronously acquired by the temperature and humidity sensor, performs matrix stitching through the feature matrix construction module to complete the preparation of the high-dimensional input feature matrix. The two-layer machine learning calibration module includes a multi-output random forest backbone module, with the output of the feature matrix construction module connected to the input of the multi-output random forest backbone module; the error output generated by the two-layer machine learning calibration module is connected to the input of the output calibration module; the SMPS connection module is connected to the input of the module training and benchmark comparison module, and the output of the module training and benchmark comparison module is connected to the output calibration module; for core aerosol modes with particle sizes ranging from 10nm to 300nm, the system utilizes a multi-output random forest to simultaneously predict multiple particle size channels to maintain the statistical correlation and spectral continuity between different particle size channels; simultaneously, the system... The system performs local reconstruction of the patch subnetwork for the extreme channels at both ends. Specifically, for the 10nm nucleation channel, it calls the independent random forest submodel with the maximum tree depth to suppress the noise amplification effect under low load conditions. For the 300nm large particle channel, it calls the gradient boost regression submodel with heavy regularization to filter out abnormal responses that may be caused by uneven flow field at the cavity edge, electrostatic migration deviation, or local stagnation. Finally, the system performs corresponding position coverage fusion of the normal channel concentration matrix predicted by the backbone model and the extreme channel concentration matrix predicted by the patch subnetwork to generate and output high-precision particle number and size distribution calibration data.

Images