Field testing of low-cost particulate matter sensors for Digital Twin applications in nanomanufacturing processes

The EU-project ASINA is testing Low-Cost Particulate Matter Sensors (LCPMS) for industrial monitoring of the concentration of airborne particles, with the purpose of integrating this sensor technology within the data collection layer of Digital Twins (DTs) for manufacturing. This paper shows the results of field performance evaluations carried out with five LCPMS from different manufacturers (Alphasense OPC-N3, Plantower 9003, Sensirion SPS30, Sensirion SEN55 and Tera Sensor NetxPM), during several field sampling campaigns, conducted in four pre-commercial and commercial pilot lines (PLs) that manufacture nano-enabled products, belonging to the ASINA and OASIS H2020 EU-projects [2,28]. Field tests consisted of deploying LCPMS in manufacturing process, measuring in parallel with collocated reference and informative instruments (OPS TSI 3330/CPC TSI 3007), to enable intercomparison. The results show the complexity and differential response of the LCPMS depending on the characteristics of the monitored scenario (PL). Overall, they exhibit uneven precision and linearity and significant bias, so their use in industrial digital systems without proper calibration can lead to uncertain and biased measurements. In this sense, simple linear models are not able to capture the complexity of the problem (non-linear systems) and advanced calibration schemes (e.g. based on machine learning), applied “scenario by scenario” and in operating conditions as close as possible to the final application, are suggested to achieve reliable measurements with the LCPMS.


Introduction and motivation
The Chemicals Strategy for Sustainability [13] emphasizes the need to deploy new production processes and technologies to enable the European chemical industry to manufacture new safe and sustainable chemicals and materials, ensuring the transition to climate neutrality.
Digital technologies are considered drivers of innovation in industrial production to reach sustainable manufacturing [8,10,13,14].The Strategic Research and Innovation Plan for safe and sustainable Chemicals and Materials (SRIP) [11] identifies the digitalization of processes to enable smart production plants and, in particular, the use of Digital Twins (DT) for the design and real-time optimization of manufacturing processes, as relevant areas of research and innovation to achieve clean, green and efficient production processes.
Digital Twins (DT) is one of the five enabling technologies supporting the new concept of Industry 5.0 [12,34].ISO/TR 24464:2020 [17] provides a short and simple definition of the DT concept, as a compound model composed of a physical asset (economic resource which exist in the real world), an avatar (digital replica of the physical asset) and an interface (to synchronize and transfer information between them).In practical terms, a DT for manufacturing is a datadriven software and hardware emulation platform, which is a digital replica of the physical asset (e.g. a manufacturing process, in our case), and includes real-time computing, real-time control, and real-time communication [1,29,33,34].Different digital technologies are used for the development and operation of the DT, such as data analytics and big data, artificial intelligence, cloud computing, internet of things, virtual reality or blockchain, among other.
DT technology is aligned with the guiding principles of (re)design proposed by Recommendation (UE) 2022/2510 [9], establishing a European assessment framework for 'safe and sustainable by design' chemicals and materials.In this regard, DT can be implemented in the (re)design of (nano)processes, with the aim of promoting material and energy efficiency (consumption reduction) and minimizing the use of chemical products and hazardous materials, through optimization of the process itself.This optimization leads in parallel to preventing and avoiding the generation of waste, hazardous by-products, and emissions.
The EU-project ASINA [2] investigates the use of DT technology for nanoprocesses optimization, with the aim of achieving safer and more sustainable nanomanufacturing processes by design, which reduce consumption of nanoforms (NFs) and energy and minimize NOAA air emissions and derived occupational exposures [3,6,26].DTs for manufacturing are complex, datadriven digital systems that require a robust data capture layer of the manufacturing process [29].In this context, ASINA is testing Low-Cost Particulate Matter Sensors (LCPMS) for the industrial monitoring of the concentration of airborne particles, with the purpose of integrating this sensor technology within the data collection layer of the DT for manufacturing [26].
LCPMS is an emerging and rapidly expanding technology used for informational and supplemental non-regulatory air quality outdoor and indoor monitoring applications [7,30,35].The term "low cost" refers to the price of the sensor compared to the price of the currently used reference instrument.Thus, LCPMS are defined by CEN/TC 137 [7], as "an Original Equipment Manufacturer (OEM) sensor module that determines number concentration and/or mass concentration and/or size distribution of airborne particles, based on the measurement of light scattered by particles.Sensors are considered to be low-cost if prices are at least 10 to 100 times lower than established".
LCPMS include nephelometers and optical particle counters (OPC).Most LCPMS provide cumulative Particle Mass Concentration (PMC), and some of them also, Particle Number Concentration (PNC), across a range of sizes.LCPMS sensors cannot usually detect particles smaller than 0.3 µm.Light scattering is sensitive to changes in environmental parameters (e.g.relative humidity), aerosol size distribution, and particle composition, which can drastically influence sensor accuracy [7,30,35].
The advantages offered by LCPMS, in terms of functionality, small size, easy installation, high spatiotemporal resolution, real time coverage, low-power consumption and low-cost, contrast with some limitations identified by ongoing research, such as high variability of data quality, the effect of humidity conditions, the long term drift in sensor response over time or the need of robust calibrations, to ensure data quality [5,7,15,19,23,32,35].
A large majority of studies with LCPMS focused on outdoor ambient air quality monitoring, although there are also some relevant applications in indoor air quality (offices, homes, buildings, schools, hospitals) [7,31,32,36].However, indoor applications in the industrial field (manufacturing processes/workplaces) are very limited but show great potential.[19,27,30].
The current availability of cost-effective digital technologies allows the construction and deployment of sensor networks based on LCPMS nodes (fixed/mobile, wireless/wired) that can be deployed in industrial processes/working areas to real-time monitor PM, providing a dense spatiotemporal resolution [31,32].The data collected by these LCPMS monitoring networks can feed the process management itself (e.g.digital prediction and optimization) and be the basis of advanced risk mapping management tools (emissions, exposures), which allow real-time management of PM risk, based on specific metrics and KPIs [26].
There are very significant differences between the typology and characteristics of the aerosols found indoors and outdoors, as well as their corresponding environmental conditions [7].Industrial aerosols are polydisperse, highly heterogeneous, and fluctuating over time, with relevant and highly variable background concentrations.Each industrial environment is unique in terms of aerosol typology, characteristics, and measurement conditions, and consequently, the performance of LCPMS can vary drastically between different types of industrial settings.
LCPMS are pre-calibrated by the manufacturer.However, calibration aerosols used in the laboratory differ considerably from aerosols found in real scenarios; so, LCPMS require additional field calibrations before implementation to ensure measurement accuracy.Postprocessing calibration strategies to improve LCPMS performance include models with corrections for hygroscopicity, traditional simple or multiple linear regression models, and algorithms based on machine learning, where the simplicity and applicability of the model is always prioritized as basic selection criterion (parsimonious approach) [5,15,16,18,20,23,24].
The absence of a standardized evaluation/calibration methodology to compare LCPMS and describe its performance, leads to the use of a variety of metrics and test procedures by the scientific community [7,15,22,23,24], which limits the inter-comparison of results and constitutes a barrier for the expansion of the technology.In this context, CEN/TC 137 is currently developing a future standard to provide guidelines and testing procedures for LCPMS to be implemented in workplaces for measuring engineered NOAA [7].
This paper shows the results of field assessments carried out with five LCPMS from different manufacturers (Alphasense OPC-N3, Plantower 9003, Sensirion SPS30, Sensirion SEN55 and Tera Sensor NetxPM) in four pre-commercial and commercial pilot lines (PLs) that manufacture nano-enabled products (PL0, PL4, PL5, PL6), belonging to the ASINA and OASIS H2020 projects [2,28].Here, we assess the performance of LCPMS comparing the response of these sensors vs. the response of two collocated reference (OPS TSI 3330) and informative (CPC TSO 3007) instruments, considering two metrics -Particle Number Concentration (PNC) and Particle Mass Concentration (PMC) -and three common size fractions for each metric (PM2.5, PM4 and PM10).The work focuses exclusively on showing the results obtained in the field tests performed with the LCPMS.The provision of calibration schemes to improve this data, as well as the strategies and algorithms to be implemented in the DT, to use this information as a proxy in NOAA airborne monitoring, are the subject of ongoing research and, therefore, they are out of the scope of this work.

Field scenarios
The main characteristics of the four pre-commercial and commercial PLs [2,4,21,25,28] selected as scenarios for the field tests with LCPMS are summarized in Table 1.PL0 -Spray coating line.PL0 is an automatic continuous commercial production line for spray coating of textile and polymeric substrates with nanocoatings (TiO2, SiO2).The manufacturing line consists of four modules: 1) Plasma, 2) Preheating, 3) Spraying (Hotspot) and finally 4) Drying.PL0 is accommodated in an industrial building and uses a centralized Local Exhaust Ventilation (LEV) system to control a large part of process emissions.
PL4 -Buckypapers pilot line.PL4 is a semi-industrial continuous production line that manufactures buckypapers through a wet dynamic vacuum filtration process.PL4 is fed from a tank by a stable dispersion of MWCNTs, manually prepared by dilution of the concentrated primary dispersion of CNTs received from the manufacturer.The pilot line consists of two main manufacturing steps: 1) Filter module (Hotspot) and 2) Washing/drying module.PL4 is accommodated in a dedicated room, equipped with a centralized LEV system.
PL5 -Doped veils pilot line.PL5 is a semi-industrial continuous production line that manufactures CNT-doped veils through a proprietary extrusion/blow process.PL5 is manually loaded with CNT-doped pellets produced from a mixture of raw CNT masterbatch and neat polymer, in a contiguous extrusion line.PL5 consists of the following stages: 1) Drying, 2) Manual loading, 3) Plasticizing, 4) Extrusion/Shaping (Hotspot) and 5) Rolling up.PL5 is accommodated in a dedicated room and equipped with a mobile LEV system located next to the Extrusion/Shaping stage.
PL6 -FXply treated prepregs pilot line.PL6 is an industrial production line that manufactures CNT treated prepregs, by combining commercial prepregs rolls and a tailored CNT-thermoplastic formulation in powder form.PL6 is an automatic manufacturing line with manual loading (prepreg roll) /unloading (treated prepreg roll).PL6 deploys the following steps: 1) Unrolling, 2) Powder Scattering, 3) Compaction, 4) Cooling and 5) Rolling up.PL6 is housed in an industrial building, where it coexists with other industrial activities.Airborne particle emissions from the scattering module are controlled by an enclosure equipped with a dust extraction system with HEPA filtration.

LCPMS -experimental setup and reference instrumentation
LCPMS and technological node.Five LCPMS units from different manufacturers [Alphasense OPC-N3, Plantower 9003, Sensirion SPS30, Sensirion SEN55 and Tera Sensor NetxPM (one unit per model)], representing the current state-of-the-art of this technology, were assessed in this work (Table 2).
A technological node (LCPMS-TN) was built on a Bopla IP66 compact electrical box (271x170x120 mm) and the five LCPMS were installed on the surface of the box, together with a temperature/relative humidity sensor (Sensirion SHT40).Inside the plastic enclosure, an ultracompact industrial PC Beckhoff C6015 was housed, to which all the sensors were connected.As each manufacturer provides its own LCPMS management software, all proprietary software was loaded on the PC and used for real-time data capture and display.The LCPMS-TN was connected to a laptop using an RJ45 Cat. 5 network cable, and the remote desktop configuration of Windows 10 was used to manage the remote operation of the LCPMS-TN as well as the download of the data initially stored on the PC, for its subsequent exploitation.
Reference and informative measurement instruments.For LCPMS intercomparison, an Optical Particle Sizer (OPS TSI 3330), widely used in evaluating the performance of LCPMS in laboratory and field tests, was used as a standard reference instrument.The OPS was used to measure PNC in the different particle size fractions selected by this work (PM2.5, PM4, PM10), also providing the corresponding mass concentrations (PMC).It is important to note that the OPS was used in this study with the original calibration parameters set by the manufacturer.
In addition to OPS, a Condensation Particle Counter (CPC TSI 3007) was also used in field testing with LCPMS, solely as an informative instrument, to capture data on the fraction of the aerosol below the measurement range of the LCPMS (<0.3 μm).Both instruments with their valid calibration periods.a The CPC was not used for intercomparison with the LCPMS, but only to capture information on the fraction of aerosol with a particle size less than 0.3 μm that cannot be detected by the LCPMS.

Measurement strategy.
The field tests consisted of deploying the LCPMS-TN with the five LCPMS in the hotspots of the PLs (Table 1) and measuring in parallel with co-located reference/informative instruments (OPS/CPC), throughout the entire work cycle of the pilot line.The duration of these cycles was different depending on the PL investigated and ranged between 2 and 4 h.
The LCPMS-TN was installed on a mobile tripod and placed at a maximum distance of 100 cm from the pilot line hotspot, and a measurement height of 150 cm above ground level, according to the average height of the worker's breathing zone.The reference and informative instruments (OPS/CPC) were installed on a transport trolley to facilitate their movement.In some cases (PL5/PL6), the LCPMS-TN and the trolley moved between various points around the hotspot to ensure multipoint measurement.Inlets were collocated together with the LCPMS-TN using extension tubes.The LCPMS-TN measured only with the Sensirion SP30 in PL0 and with all the LCPMS in the rest of the PLs.
Here, we have focused our work on measuring the most common particle sizes ranges used by environmental and occupational conventions, such as PM2.5 (particles with diameter ≤ 2.5 μm), PM4 (d ≤ 4 μm) and PM10 (d ≤ 10 μm), both for PNC and PMC (see Table 2).

Data treatment and analysis
Time series collected by LCPMS and reference instruments, initially sampled at 1 s, were time-paired, filled in for missing data, outliers removed, and after spectrograms analysis, resampled to 5 s, to reduce the amount of data without losing relevant signal information.
This data pre-treatment as well as the subsequent statistical analysis were carried out with Python and Matlab software.Time series graphs, scatterplots and descriptive statistics were used to explore the general patterns of data collected by LCPMS and reference/informative instruments and enable intercomparison of results and performance analysis, for the selected metrics and size fractions.
The linear regression (LR) model, traditionally used for intercomparison studies, was selected in this work as the simplest approximation to explore the performance of LCPMS and data quality.The slope (S) of the linear regression and the Y-intercept (Y-i); the coefficient of determination R 2 and the Root Mean Square Error (RMSE) were the single-value metrics used to analyze the bias, linearity, and error of the LCPMS vs. the reference OPS.

General response patterns of LCPMS
Figure 1 shows the signals measured in the four PLs by the informative CPC, in the 0.01-1 μm range (blue line), and the OPS in the channel PM2.5 (red line).The differences between both signals are evident and suggest that a relevant fraction of the aerosol below a particle size of 0.3 μm cannot be detected by the OPS and consequently neither by the deployed LCPMS.The concentrations reported by the CPC in the PLs ranged from 7346 #/cm 3 in PL6 (the cleanest setting) to 51940 #/cm 3 in PL6, where the highest PNC values were recorded.

PL0 PL4 PL5 PL6
Figures 2 and 3 show respectively the time series of the LCPMS and the OPS reference sensor, and the scatter plots of LCPMS vs. the OPS reference instrument, for the two metrics (PNC and PMC, sampling rate 5 s), the three particle sizes (PM2.5, PM4 and PM10) and the four pilot lines (PL0, PL4, PL5 and PL6) considered in this work.Both graphs show the complexity and the relevant differences in the response of the LCPMS, depending on the monitored scenario.For example, the signal recorded by the LCPMS and the OPS in PL1 shows a repetitive periodic pulsing signal, associated with gun projections in a spray chamber; PL4 and PL5 show complex signals with significant peaks and valleys, corresponding to the "start-up/production/stop" cycles of these pilot lines, and finally, PL6 shows a process signal with a dominant noise component.The LCPMS responses (Figure 2) mimic the OPS reference monitor response shape in all channels but show a different amplitude value depending on the size fraction and the investigated scenario.In particular, the response of the PMC10 channel contains a noise component that notably dirty the signal.
Figure 3 intuitively separates two different LCPMS performances that are highly dependent on the monitored scenario.On one side, scatterplots of pilot lines PL4 and PL5 (scenarios with the highest mean particle concentration values), show low-moderate dispersion of data points and moderate-good agreement between the LCPMS and the reference monitor.However, the scatterplots of the pilot lines PL0 and PL6 (scenarios with lower particle concentration), show a very high scattering of the data, no linear trend and therefore a poor or null fit of LCPMS to the linear model.Figure 3 also shows that the dispersion of the data increases significantly as the size fraction increases.

LCPMS performance
Table 3 shows the basic descriptive statistics of the time series monitored by the LCPMS and the reference and informative instruments (OPS/CPC).In addition, Table 4 shows the results of the metrics used for the analysis of the bias, linearity, and error of the LCPMS, by comparing the concentrations recorded by these sensors and the reference instrument (OPS), for the different scenarios (PL0 to PL6), metrics (PNC, PMC) and size fractions considered (PM2.5, PM4, PM10).The statistical parameters show evident differences in the performance of the LCPMS, depending on the metric, size fraction and especially the investigated scenario (PL).

Particle Number Concentration (PNC)
The mean PNC reported in the PLs by the group of LCPMS range between 25.80 and 570.21 #/cm 3 , for the set of size fractions (Table 3).The average values of PNC in PL0 and PL6 are one order of magnitude lower than those measured in PL4 and PL5.In addition, PL4 and PL5 exhibit the highest mean values of PNC.Sensiron SPS30 and Tera Sensor NextPM sensors always provide mean PNC values higher than the values provided by the reference sensor (OPS), for all size fractions and PLs.In addition, the Tera Sensor NextPM sensor always records higher values than the Sensiron SPS30 sensor.
The differences between the PNC measured by the LCPMS and by the OPS do not change with the increase in particle size, since the value of the PNC2.5 fraction is dominant and largely influences the results of the other two fractions, PNC4 and PNC10.
The results in Table 4 show that for PNC, the scale error represented by the slope of the regression line (S), ranges between 0.29 and 1.92 for the set of LCPMS and PLs.LCPMS generally overestimate the measurement with respect to the reference monitor at PL4 and PL5 (S>1) and underestimate it at PL0 and PL6 (S<1).The displacement error characterized by the value of Y-intercept (Y-i) is very significant in all cases and ranges between -55.58 and 81.44 #/cm 3 .
The coefficient of determination R 2 ranges from 0.11to 0.97.The LCPMS show high values of R 2 and high linear correlation between PNCs measured by LCPMS and the reference sensor (OPS) in pilot lines PL4 and PL5 (0.72-0.97), lower values in PL6 (around 0.40) and values close to 0 in PL0 (0.11).
The measurement error (RMSE) ranges between 4.42 and 687.18 #/cm 3 and the highest values are found in the pilot lines PL4 and PL5.Bias, linearity, and error do not change with the increase in particle size for the set of LCPMS and PLs.

Particle Mass Concentration (PMC)
The mean PMC reported by the LCPMS range between 2.72 and 124.46 μg/m 3 for the set of size fractions: PL0 (7.01-7.67 μg/m 3 ), PL4 (11.62-89.90μg/m 3 ), PL5 (6.32-124.46μg/m 3 ) and PL6 (2.72-23.67μg/m 3 ).The highest mean values were reported in PL5 (124.46 μg/m 3 for PMC10).The mean values of PMC2.5 provided by the LCPMS in PL0 and PL6, are similar to or slightly higher than the mean values provided by the reference instrument, while for PMC4 and PMC10 the values are significantly lower, especially for this last fraction of size.In PL4 and PL5, the mean PMC values provided by the LCPMS are always much higher than those of the reference instrument, for all size fractions and all LCPMS, except for the Alphasense OPC-N3 sensor.This sensor shows a differential behavior with respect to the rest of LCPMS, since the values it reports are always lower (20 to 30%) than those of the reference instrument, both for PMC2.5 and for PMC10.
In PL4 and PL5, the differences between PMC provided by the LCPMS and the OPS generally decrease as the size fraction increases.However, in PL0 and PL6, these differences increase significantly with particle size, especially for PMC10.In this fraction, the differences between the concentrations reported by the OPS and the LCPMS are in the range of 10-20 μg/m 3 .All this suggests that there are coarse particles measured by the OPS that have not been detected by the LCPMS.Once again, the Alphasense OPC-N3 sensor moves away from this behavior in PL6, since it presents PMC10 values similar to the value of the reference instrument.The Plantower 9003 sensor is generally the one that reports the highest mean values and is furthest from the reference OPS sensor, for all size ranges.
The scale error shows in Table 4 for PMC, ranges between 0 and 9.59 for the set of LCPMS and PLs.In PL4 and PL5, this significant bias is generally much greater than 1 (1.83-9.59)and consequently the PMC is severely overestimate by the LCPMS.This overestimation decreases as the size of fraction increases.Conversely, in PL0 and PL6 the bias is always less than 1 (the slope value is close to zero or is 0 for all fractions) and the PMC is underestimated by the LCPMS.In this case, the underestimation grows with the increase of the size fraction.However, the Alphasense OPC-N3 sensor deviates from the patterns described above and presents a differential behavior with respect to the rest of the LCPMS, since it underestimates the PMC in all the PLs (0.05-0.46), and this underestimation increases with the size of particle.The maximum scale error corresponds in general with the Plantower 9003 sensor and the smallest error with the Alphasense OPC-N3 sensor.The behavior of the two sensors Sensirion SPS30 and Sensirion SEN55 is similar.
The displacement error ranges in the interval [-68.07/22.44μg/m 3 ].This bias is positive in PL0 and PL6 and negative in PL4 and PL5.The Alphasense OPC-N3 sensor also presents a different pattern here, as the value provided for Y-i is always positive across all PLs and size fractions.Furthermore, this LCPMS is the one with the lowest displacement error for all the PLs.This bias increases with the increase in particle size, except for PL0, where it decreases slightly.The largest value of bias is generally provided by the Plantower 9003 sensor, followed by the Sensirion SPS30 and SensirionSEN55 sensors.
PMC2.5 exhibits a very significant coefficient of determination (R 2 ) in PL4 and PL5, for the whole group of LCPMS [PL4 (0.68-0.82) and PL5 (0.87-0.98)], indicating a strong correlation between the measures provided by the LCPMS and the data recorded by the OPS reference sensor.However, R 2 is 0 or practically 0 in PL0 and PL6 for all the size fractions, showing the null linear correlation between the data from the LCPMS and the reference monitor in these scenarios.Furthermore, the results also show that R 2 decreases significantly as the fraction size increases, for the group of LCPMS.This decrease in R 2 is especially significant for the Alphasense OPC-N3 sensor in PL4 and PL5, where R 2 moves from high values for PMC2.5 (0.7-0.87) to practically zero values (0.05-0.08) for PMC10 fraction.
The measurement error quantified by the RMSE parameter ranges between 0.85 and 169.8 μg/m 3 , depending on the PL and/or the fraction considered, and grows with the particle size.The Alphasense OPC-N3 sensor generally exhibits the lowest RMSE values.This pattern is especially pronounced in PL4 and PL5, where this sensor presents concentration differences of up to an order of magnitude with the rest of LCPMS.

Limitations
It is important to note that this study has several spatiotemporal and technological limitations that need to be considered to properly interpret the representativeness of the results: 1) The sample of industrial processes investigated with the LCPMS is small and limited to four PLs.Therefore, these measurements are only representative of the monitored processes and the manufacturing and sampling conditions established in each case for field tests; 2) The measurements with the LCPMS in the PLs were carried out in short-term manufacturing cycles (2-4 h), specially enabled for the measurement campaigns; 3) The LCPMS were tested in each monitored process in one specific single location (hotspot); 4) Only one unit from each of the five LCPMS was tested by this study; and 5) LCPMS use proprietary algorithms and internal configuration parameters that, with few exceptions, are unknown to the end user.This introduces an additional bias between LCPMS that makes intercomparison of results difficult.

Conclusions and future work
This work shows the results of field tests carried out with five LCPMS from different manufacturers (Alphasense OPC-N3, Plantower 9003, Sensirion SPS30, Sensirion SEN55 and Tera Sensor NetxPM), in four pre-commercial and commercial pilot lines.
The time series collected by the LCPMs show the great complexity and differential response of these sensors, depending on the typology and characteristics of the industrial monitored scenario.Even for the same investigated pilot line, differential responses of some sensors with respect to the group of LCPMS have been observed (the Alphasense OPC-N3 sensor, in particular).
The results separate two different LCPMS performances, corresponding to the pilot lines PL4 and PL5 on the one hand (scenarios with the highest mean particle concentration values) and PL0 and PL6 on the other (scenarios with lower particle concentration).There is a moderate-good agreement between the LCPMS and the reference monitor in PL4 and PL5 (high R 2 values), while the agreement in PL0 and PL6 is very poor or non-existent (very low or practically null R 2 values).LCPMS generally overestimate reference concentrations (OPS) at PL4 and PL5 and underestimate them at PL0 and PL6.These deviations grow as the size of the particle fraction increases, In general, LCPMS exhibit uneven precision and linearity and significant bias, so their use in industrial digital systems without proper calibration can lead to uncertain and biased measurements.
In this sense, the results also show that the simple linear model traditionaly used to characterize and calibrate LCPMS, is not able to capture all the complexity of the problem in the PLs (nonlinear systems).Thus, advanced calibration schemes (e.g. based on machine learning), applied "scenario by scenario" and in operating conditions as close as possible to the final application, are suggested to achieve reliable measurements with the LCPMS in industrial scenarios.
This calibration approach, together with the definition of strategies and algorithms to use LCPMS data as a proxy in airborne NOAA monitoring, are the subject of ongoing research in project ASINA.

Figure 1 .
Figure 1.Time series of particle number concentration (PNC) recorded by the reference and informative instruments in the PLs: OPS TSI 3330 (red line) and CPC TSI 3007 (blue line).The graphs show the response of both instruments to the signal produced by the industrial process, in the 0.01-1 μm range for the CPC (blue line) and in the PM2.5 channel for the OPS (red line).

Table 2 .
LCPMS, reference/informative instruments, metrics and size fractions used by LCPMS in this study.

Table 4 .
LCPMS performance metrics for concentrations (PNC, PMC) and particle size fractions (PM2.5, PM4 and PM10) considered in this study.Each LCPMS channel has been compared to the corresponding channel of the reference instrument (OPS TSI 3330).Abbreviations are the same as in Table3, except for: S, Slope; I; Intercept; R 2 , Coefficient of determination and RMSE, Root Mean Square Error.
The highest RMSE values are generally associated with the Plantower 9003 sensor.