LED array-based multi-angle light scattering for aspirating smoke detection and classification

Abstract

Although smoke detectors are actively being studied to reduce false fire alarms, they still face challenging issues such as complex and elaborate alignment, high cost, large size, and poor performance. In particular, most smoke detection systems based on Mie scattering, which rely on single-scattering measurements, may not perform effectively in real-world environments where multiple scattering occurs. We present an advanced smoke detection instrument for aspirating smoke detection and classification based on multiple scattering. Multi-angle light scattering with an LED array instead of angle-positioned PDs was measured, and the unique optical property ratios of fire and non-fire aerosols were calculated. The feasibility of smoke detection and classification was verified by evaluating the classification performance of 10 types of fire and non-fire aerosols using general supervised learning algorithms. The advanced smoke detection instrument features a simple design, making it cost-effective and compact. In addition to reducing false fire alarms, it is expected to contribute to choosing appropriate fire extinguishers based on fire class and advancing research of complex fire detection.

Introduction

Smoke detectors are one of the most practical fire detectors and are widely utilized due to their simple design, compact size, low cost and high performance. Installation of smoke detectors is mandatory in many countries such as South Korea, the United States, Canada, the United Kingdom, Australia, and New Zealand. Photoelectric smoke detectors, which respond more quickly to smoldering fires, are generally preferred over ionization smoke detectors^1,2. Photoelectric smoke detectors, which operate on the simple principle of illuminating smoke with light from a light-emitting diode (LED) and measuring the light scattered by the smoke using a photodetector (PD)³, have an inherent drawback: they respond indiscriminately to all aerosols (e.g. fine dust, water vapor, cooking smoke, and cigarette smoke) and trigger an alarm. These perpetual nuisance alarms result in a waste of firefighting manpower and resources^4,5 and cause people to turn off smoke detectors, which can even lead to significant property damage from actual fires.

To reduce false alarms, a variety of smoke detection methods have been proposed, such as using multi-wavelength light sources^{6, 7, 8–9}, detecting multi-angle light scattering (MALS)^{10, 11, 12–13}, measuring polarization^{14, 15, 16–17}, and utilizing extra sensors^{18, 19–20}. Among these methods, MALS, a technique that obtains light scattered by a sample at a plurality of angles, enables the determination of particle size^{21, 22–23} and molar mass^24,25, and is therefore utilized to investigate fire and non-fire aerosols²⁶ as well as fine particles²⁷ and dust²⁸. In addition, various approaches for upgrading MALS systems have been introduced, such as improving computational speed by using low-complexity Mie scattering algorithms instead of high-complexity signal processing methods²⁹, and enabling real-time analysis of particle morphology and size distribution of combustion-generated aerosols³⁰. However, the commercialization of conventional MALS-based smoke detection is highly challenging due to practical issues such as complex and elaborate alignment, high costs, large size, and poor performance. In particular, most smoke detection systems based on Mie scattering, which rely on single-scattering measurements, may not perform effectively in real-world environments where multiple scattering occurs³¹.

An aspirating smoke detector (ASD), which draws smoke into a detection chamber through a pipe, detects smoke based on a principle analogous to that of a photoelectric smoke detector. The ASD employs a laser diode instead of an LED and is capable of monitoring low concentrations of smoke quickly^32,33. Like photoelectric smoke detectors, ASDs respond to all aerosols and are vulnerable to false alarms, and are therefore mainly used in confined spaces such as clean rooms or energy storage systems. Research on ASDs has not been actively pursued to explore new approaches, improve performance, or reduce false alarms.

[See PDF for image]

Fig. 1

Configuration for smoke detection of ASDI. Smoke detection uses a 1D LED array, single lens, and two photodetectors (PDs). The LEDs are driven by sequentially turning on and off. During this process, LED1 and LED2 are wired in parallel to operate simultaneously; the same applies to LED5 and LED6. PD1 monitors the light intensity of the LEDs, and PD2 measures the scattered and absorbed light signals from smoke. denotes the illumination angle.

Here, we present an advanced smoke detection instrument (ASDI) that detects smoke through air sampling and classifies various types of smoke, specifically fire and non-fire. ASDI measures multi-angle light scattering with an LED array instead of angle-positioned PDs and classifies five types of fire and five types of non-fire aerosols via the ratio of scattering coefficients to absorption coefficients calculated based on multiple scattering effects to reduce false alarms.

Results

LED array-based MALS

Conventional MALS acquires scattered light by positioning multiple PDs at different angles or by moving a single PD along the scattering angle, whereas we measure scattered light at different angles with a one-dimensional (1D) LED array and two stationary PDs. The configuration is depicted in Fig. 1.

A 1D LED array with 6 LEDs was employed as a light source. The light from the 1D LED array is collimated via a lens and projects onto smoke particles. The light scattered and absorbed by the smoke particles is detected by PD2. In detail, for LED1, LED2, LED5, and LED6, the light scattered, refracted, and diffracted by the smoke particles is collected, while the light from LED3 and LED4 passes through the smoke particles, and the transmitted light is measured. PD1 records the light intensity of the LEDs to compensate for their brightness, which depends on ambient temperature³⁴ and self-heating³⁵. The temperature-dependent variations in LED light intensity were effectively compensated by PD1 (Fig. S1). The 1D LED array is sequentially turned on and off, and PD1 and PD2 record the light signal each time. The six LEDs are not controlled individually, but LED1 and LED2 as well as LED5 and LED6 are driven simultaneously. In one cycle, the light signals are sequentially obtained four times: the first measurement is for LED1 and LED2, the second for LED3, the third for LED4, and the fourth for LED5 and LED6. The measurement time for one cycle is 15.38 ms, corresponding to a sampling rate of 65 Hz. LED1 to LED6 are respectively positioned at a distance of +7 mm, +5 mm, +1 mm, mm, mm, and mm from the center, and the corresponding illumination angles () of each LED are +25.02 degrees, +18.43 degrees, +3.81 degrees, degrees, degrees, and degrees, respectively. The distance between the 1D LED array and the lens is 15 mm, and the detector active area measures 4 mm 5 mm. Ultraviolet LEDs, which are sensitive to aerosols with small particle sizes, and near-infrared LEDs, which offer high PD responsiveness and are more suitable for detecting relatively larger particles, were employed as light sources.

Signal processing for obscuration and optical properties

To alert for fire or non-fire events, ASDI processes signals in the following sequence. First, the obscuration (OBS) and optical properties (i.e., scattering and absorption coefficients) are monitored at 1-second intervals. When the OBS exceeds 1%/m, optical property data begins to accumulate. Once the OBS reaches 15%/m, the accumulated optical properties are used to calculate the optical property ratios ( to ), which are then classified using the trained machine learning model to determine whether the event is fire or non-fire. The signal processing process is exhibited in Fig. 2. The raw data measured at a sampling rate of 65 Hz is averaged at 1-s intervals, and the light signal obtained from the white LED (i.e., LED4) can be expressed in OBS as

1

[See PDF for image]

Fig. 2

Signal processing for OBS and optical properties. (a) OBS calculated from the raw data of LED4. (b) Measured relative scattering coefficients at 450 nm (blue solid line) and 940 nm (red solid line). (c) Measured relative absorption coefficients at 940 nm (red solid line) and in the visible wavelength range (blue solid line). (d) Relative scattering coefficient versus relative absorption coefficient at 940 nm (blue circles); relative scattering coefficient in the visible range versus relative absorption coefficient at 450 nm (red circles); relative scattering coefficient at 450 nm versus relative absorption coefficient at 940 nm (yellow circles); and relative scattering coefficient at 940 nm versus relative absorption coefficient in the visible range (violet circles). (e) Relative absorption coefficient at 940 nm versus that in the visible range. (f) Relative scattering coefficient at 940 nm versus that at 450 nm.

where I is the measured light intensity, is the initial light intensity, and d is the path length, which is similar to the OBS measurement of the white light obscuration system of Underwriters Laboratories³⁶. Figure 2a shows the measured OBS as a function of time. The measured OBS is calibrated to the smoke concentration detected by sensitivity test equipment (Haeyoung Precision, SIT-250-CHAMBER) used for smoke detector performance testing, and detailed experimental methods are described in section S2 of Supplementary Information. The limit of detection (LoD) of OBS was found to be approximately 1.65%/m, which was calculated as where is the standard deviation of OBS measured in ambient air and S is the slope, which is 1 in this case.

Scattering coefficient () and absorption coefficient () were computed using direct measurements with the total transmission (TT) and collimated transmission (CT) as suggested by Hohmann et al.³⁷. For thick samples with strong scattering, the scattering and absorption coefficients can be obtained by simplified equations:

2

where TT is the total transmittance measured from LED1 to LED3 or from LED4 to LED6, and CT is the collimated transmittance emitted from LED3 or LED4. Fire and non-fire aerosols have individual physical properties such as particle size, shape, and refractive index (RI)^21,36,38,39, and the scattering and absorption coefficients are determined by these unique properties^{40, 41–42}.

In this paper, ASDI measures the relative scattering coefficients at wavelengths of 940 nm and 450 nm (Fig. 2b), the relative absorption coefficients at wavelengths of 940 nm and in the visible range (444 nm to 641 nm) (Fig. 2c). Measured relative scattering and absorption coefficients are accumulated over an OBS range of 1%/m to OBS of 15%/m, and six ratios ( to ) are calculated using first-order polynomial fitting of the accumulated data (Fig. 2d–f), which serve as unique fingerprints of aerosols. represents the ratio of the absorption coefficient to the scattering coefficient at a wavelength of 940 nm. represents the ratio of the absorption coefficient in the visible wavelength range to the scattering coefficient at 450 nm.

[See PDF for image]

Fig. 3

Sample preparation for optical property measurement. (a) Samples fabricated with PS beads of different particle sizes (100 nm, 200 nm, 500 nm) at concentrations ranging from 0 to 0.1 wt%. (b) Samples with various RIs prepared using PS, PMMA, and carbon particles. Samples with PS and PMMA particles were fabricated at concentrations ranging from 0 to 0.1 wt%, and samples with carbon particles were fabricated at concentrations ranging from 0 to 0.01 wt%. (c) Optical thickness of PS beads, PMMA beads, and carbon nanoparticle samples measured at 940 nm for various concentrations.

represents the ratio of the absorption coefficient at 940 nm to the scattering coefficient at 450 nm. represents the ratio of the absorption coefficient in the visible wavelength range to the scattering coefficient at 940 nm. represents the ratio of the absorption coefficient in the visible wavelength range to the absorption coefficient at 940 nm. represents the ratio of the scattering coefficient at 450 nm to the scattering coefficient at 940 nm.

Validation of optical property measurements for prepared samples

In Lambert–Beer law (), the scattering regime can be classified into single scattering (), multiple scattering (), and diffusion (), where I is the measured light intensity, is the initial light intensity, and is the optical thickness⁴³. To assess the optical coefficient measurements in environments where single and multiple scattering occur, samples were prepared at different scattering regimes.

Although models that directly estimate scattering and absorption coefficients perform well when scattering is stronger than absorption, this study focused on smoke detection and classification rather than on accurately determining the values of scattering and absorption coefficients.

To verify the feasibility of categorizing a variety of smokes based on particle size and RI in both single and multiple scattering regimes, smoke-mimicking samples were fabricated. For samples with different particle sizes, polystyrene (PS) beads with diameters of 100 nm, 200 nm, 500 nm, and 1000 nm were mixed with polydimethylsiloxane (PDMS) and solidified to prepare concentrations from 0 to 0.1 wt%. Figure 3a represents images of samples fabricated with PS beads of different particle sizes, mimicking white smoke. Samples with various RIs were fabricated by mixing PS beads, polymethyl methacrylate (PMMA) beads, or carbon nanoparticles with PDMS. Samples of PS beads and PMMA beads were prepared at concentration of 0–0.1 wt%, and carbon nanoparticles samples were prepared at concentration of 0–0.01 wt% (Fig. 3b). Figure 3c presents the optical thickness of PS beads, PMMA beads, and carbon nanoparticle samples investigated at a wavelength of 940 nm. At a wavelength of 589 nm, the RIs of PS beads, PMMA beads, and PDMS are 1.5905 ± 0i, 1.4905 ± 0i, and 1.3947 ± 0i, respectively. The RI of black carbon nanoparticles is 1.67 ± 0.67i at a wavelength of 525 nm.

For all samples with different particle sizes and RIs, scattering and absorption coefficients were measured and compared using a reference instrument and ASDI. ASDI investigated the relative scattering and absorption coefficients in free space using the same configuration as Fig. 1. The reference instrument, which uses spatial low-pass filtering and an integrating sphere, measured CT and TT with high precision and calculated the scattering and absorption coefficients with the identical model formula³⁷. The experimental setup of the specific reference equipment for CT and TT measurements is described in section S3 of Supplementary Information.

To examine the effect of illumination angle on the accuracy of optical property ratio measurements, optical coefficients were investigated over a range of illumination angles. Optical coefficients were measured for samples prepared with 100 nm PS beads at different concentrations (0 to 0.1 wt%) under various illumination angle ranges, and the optical property ratio () was calculated (Fig. 4). The results showed that as the illumination angle range increased, the relative scattering coefficient became more consistent with the scattering coefficient measured by the reference instrument. This indicates that a broader illumination angle enables more accurate and sensitive estimation of optical property ratios.

[See PDF for image]

Fig. 4

(a) Relative scattering coefficients obtained by the ASDI sensor module and scattering coefficients detected by the reference instrument (purple empty circles) under various illumination angle ranges. (b) Measured as a function of illumination angle range.

[See PDF for image]

Fig. 5

Experimental results of optical property measurements for prepared samples. Scattering coefficient versus absorption coefficient of samples prepared with different particle sizes measured by (a) ASDI and (b) reference instrument. (c) measured with ASDI (black empty circles) and the reference instrument (orange empty circles) at various particle diameters (d). Scattering coefficient versus absorption coefficient of samples fabricated with various particles measured by (d) ASDI and (e) the reference instrument. (f) measured with ASDI (black empty circles) and the reference instrument (orange empty circles) from PS, PMMA and carbon particles. All samples were measured five times, and the error bars indicate the standard deviation.

[See PDF for image]

Fig. 6

ASDI for aspirating smoke detection. (a) Front and (b) side views of the customized ASDI. The blue solid arrow indicates the outlet pipe. (c) Photograph of inlet pipes with multiple perforations. The yellow solid arrows point to the inlets. (d) Photo of the 1D LED array as the light source for smoke detection. (e) Picture of the sensor module from ASDI. (f) Illustration of the internal structure of the sensor module.

The scattering and absorption coefficients of samples prepared with varying particle sizes were evaluated at different concentrations (i.e., different scattering regimes) using ASDI and a reference instrument, and the results are indicated in Fig. 5a, b, respectively. In both the graphs visualized from ASDI and the reference instrument, the scattering and absorption coefficients increased with higher concentration, and comparable profiles were observed during the transition from single-scattering to multiple-scattering regimes. To examine this relationship quantitatively, , which is the slope of the scattering coefficient with respect to the absorption coefficient at a wavelength of 940 nm, was calculated as a function of particle diameter. It was confirmed that the measured decreases proportionally with increasing particle diameter (Fig. 5c). This can be accounted for by the fact that the mean free path depends on particle size. For the same total cross-sectional area, smaller particles in the medium have a shorter mean free path compared to larger particles⁴⁴, and this increases the scattering coefficient, which is inversely proportional to the mean free path⁴⁵. It should be noted that the discrepancy between the optical properties observed by ASDI and the reference instrument arises from inaccurate TT measurements due to the limited illumination angle. The illumination angle of the LED located at the edge of the 1D LED array in ASDI is approximately 28 degrees. The scattering and absorption coefficients of samples fabricated with particles (i.e., PS, PMMA, and carbon) of different RIs at various concentrations were evaluated by ASDI and the reference instrument (Fig. 5d, e). The value estimated from the scattering and absorption coefficients of samples with different particles varies with the RI⁴⁶ and is exhibited in Fig. 5f. As previously mentioned, differences between the scattering and absorption coefficients measured by ASDI and those measured by the reference instrument were observed due to the finite illumination angle; however, the overall trends remained consistent.

ASDI for aspirating smoke detection

To evaluate the feasibility of smoke detection and classification for fire and non-fire aerosols, an ASDI was customized, and the front and side views of the ASDI are displayed in Fig. 6a and b, respectively. After passing through a mesh screen, fire and non-fire aerosols are aspirated by a fan into a pipe inlet with multiple perforations, as shown in Fig. 6c. The aspirated smoke is then detected by the sensor module of the ASDI and subsequently exhausted through an external duct. The sensor module of the ASDI, capable of monitoring LED array-based MALS, was custom-manufactured based on the configuration illustrated in Fig. 1. Photographs of the 1D LED array, which serves as the light source for the sensor module, and of the sensor module are depicted in Fig. 6d and e, respectively, and the internal structure of the sensor module is presented in Fig. 6f. Filter paper, kerosene, polyvinyl chloride (PVC), acrylonitrile butadiene styrene (ABS), and cotton wick were selected as fire sources, while water vapor, oil mist, smoke check spray, fog machine, and fly ash were selected as non-fire sources (Fig. 7). Experiments were performed on five types of fire sources and five types of non-fire aerosols. Table 1 shows the particle size distributions for the five fire sources and five non-fire aerosols^12,39,47.

[See PDF for image]

Fig. 7

Five types of fire and five types of non-fire sources. (a) Filter paper. (b) Kerosene. (c) PVC. (d) ABS. (e) Cotton wick. (f) Water vapor. (g) Oil mist. (h) Smoke check spray. (i) Fog machine. (j) Fly ash.

Table 1. Particle size distribution of fire and non-fire aerosols.

Fire aerosols	Median particle size (nm)	Standard deviation	Non-fire aerosols	Median particle size (m)	Standard deviation
Filter paper	272.2	1.69	Water vapor	9.40	1.98
Kerosene	300.8	1.70	Oil mist	1.60	2.20
PVC	263.2	1.77	Smoke check spray	7.25	1.99
ABS	175	24	Fog machine	0.25	0.05
Cotton wick	123.0	1.74	Fly ash	<45

[See PDF for image]

Fig. 8

obtained as a function of OBS from filter paper, ABS, and water vapor.

Measured optical property ratios for fire and non-fire aerosols

In order to extract optical property ratios (–) with high precision and accuracy, accumulation of data for different aerosol concentrations is required, and the accuracy and precision of ASDI are determined by the accumulated data. The measured for the accumulated concentration of representative fire and non-fire aerosols was investigated (Fig. 8) and was found to converge above 6%/m, which represents the minimum required concentration. In this paper, the six ratios are calculated by accumulating data up to 15%/m, which is the fire alarm concentration standard in South Korea.

[See PDF for image]

Fig. 9

Optical property ratios calculated for five types of fire and five types of non-fire aerosols. (a) , (b) , (c) , (d) , (e) , and (f) .

All experiments on five types of fire and five types of non-fire aerosols were performed ten times with three sensor modules to obtain six optical property ratios, and the results are portrayed in Fig. 9. Error bars indicate standard errors. Experimental results for , , , and exhibited clearly distinct patterns for each sample. However, the results for , which represents the absorption coefficient ratio between the visible wavelength and 940 nm, and , which describes the scattering coefficient ratio between 450 nm and 940 nm, depicted similar but slightly different profiles since the scattering coefficient and absorption coefficient are dependent variables, as described by the equation (, where is the extinction coefficient and is the wavelength.

The are determined by factors such as particle size and RI, which influence the scattering and absorption coefficients. A greater RI difference between the particles and the surrounding medium results in an increase in the scattering coefficient⁴⁸. The effect of particle size on the scattering coefficient requires different interpretations depending on the scattering regime. In the single scattering regime, where Rayleigh and Mie scattering dominate, the scattering coefficient increases with larger particle sizes⁴⁹. In contrast, in the multiple scattering regime, the scattering coefficient tends to increase with smaller particle sizes⁵⁰. The absorption coefficient is determined by how effectively a material absorbs light. As light passes through a medium, it may be absorbed through resonance with specific vibrational modes of molecules or through the conversion of vibrational energy into heat. Note that the absorption coefficient at a specific wavelength depends on how well the photon energy matches the vibrational energy levels of the material, making the wavelength of the light source an important factor.

Multi-class classification evaluation using supervised learning algorithms

A fire alarm system triggers an alert when the smoke concentration exceeds the specified OBS standards. A low OBS standard may lead to frequent false alarms, whereas a high OBS standard may result in delayed fire alarms, potentially exacerbating fire damage. Therefore, it is crucial to select an appropriate OBS standard that ensures accurate fire detection and rapid alarm activation.

The classification performance was evaluated using the K-Nearest Neighbors (KNN) machine learning algorithm with six optical property ratios quantified from varied OBS levels for five types of fire and five types of non-fire aerosols. The dataset consisted of 300 samples. The distance metric was set to Euclidean, with the number of nearest neighbors set to 5, and the nearest neighbor search method was set to default. 10-fold cross-validation was performed using MATLAB (MathWorks, R2024b). The results of the classification performance are presented in Table 2. For OBS concentrations of 15%/m or higher, classification performance exceeded 0.94 in all evaluation metrics (accuracy, recall, precision, F1 score). Figures 10a and b represent the confusion matrices for classifications at OBS 6%/m and 15%/m, respectively.

Table 2. Classification performance results in different OBS ranges.

OBS (%/m)	Accuracy	Recall	Precision	F1 score
6	0.873	0.873	0.877	0.873
10	0.910	0.910	0.912	0.910
15	0.943	0.943	0.945	0.943
20	0.943	0.943	0.944	0.941

[See PDF for image]

Fig. 10

Confusion matrices for classification of fire and non-fire aerosols by OBS. Classification results obtained by calculating optical property ratios (a) in the OBS range of 1%/m to 6%/m and (b) in the OBS range of 1%/m to 15%/m.

We evaluated and compared the multi-class classification performance using four machine learning algorithms: decision tree (DT), KNN, Naïve Bayes (NB), and support vector machine (SVM) (Fig. 11). The classification evaluation metrics for each algorithm are recorded in Table 3. Classification performance was found to be above 0.93 for all models except the SVM, which is more suitable for binary classification^51,52. The dataset used to train and evaluate all machine learning models consisted of 300 samples, collected from five types of fire and five types of non-fire aerosols. Each aerosol type was measured 10 times with three identical ASDI sensor modules, resulting in 300 independent measurements. Note that, to improve and stabilize the generalization performance of the model under limited exposure conditions, data were collected from three sensors instead of a single one and used for training and classification evaluation. Optical property ratios ( to ) were calculated by polynomial fitting of accumulated OBS data up to 15%/m. The dataset was further processed with label assignment for multi-class supervised learning. A 10-fold cross-validation approach was used to avoid overfitting. Training and classification were performed using MATLAB with CPU-based computation, which is well suited for small-scale datasets and conventional machine learning algorithms. The hardware system used an Intel i9-10900X processor and 128 GB of memory.

Discussion

ASDI has been demonstrated to detect smoke by aspirating air and classifies fire and non-fire aerosols via the calculation of optical property ratios that reflect multiple scattering. Smoke-mimicking samples were fabricated for single and multiple scattering conditions, and their optical coefficients were investigated in different scattering regimes, and the results were found to be similar to the optical coefficients measured by reference instrument.

Fire and non-fire aerosols exhibit diverse particle morphologies, ranging from spherical particles like water vapor to asymmetric structures such as chain-like or fractal-shaped soot. The scattering characteristics for these morphologies of the particles need to be analyzed depending on the scattering regime. In the single scattering regime, the asymmetry of particles directly affects the direction and intensity of polarization regardless of the polarization state of the incident light. Scattering matrix elements such as S12, S33, and S34 are sensitive to particle shape and depend on the measurement angle, whereas the total scattered intensity represented by S11 is less sensitive to particle morphology⁵³. In the multiple scattering regime, the scattering effects due to particle morphology differ depending on whether the incident light is polarized. Polarized incident light tends to exhibit an asymmetric scattering intensity distribution when interacting with non-spherical particles, and this asymmetry gradually diminishes as multiple scattering progresses. In contrast, unpolarized incident light may become polarized upon the first scattering by non-spherical particles, resulting in asymmetric scattering, but it rapidly becomes randomized through multiple scattering and eventually converges to isotropic behavior^54,55.

[See PDF for image]

Fig. 11

Confusion matrices for classification of fire and non-fire aerosols by various machine learning algorithms. Classification performed by machine learning algorithms of (a) DT, (b) KNN, (c) NB, and (d) SVM.

Table 3. Classification performance results with various machine learning.

Model	Accuracy	Recall	Precision	F1 score	Training time (sec)	Inference time (sec)
DT	0.953	0.953	0.956	0.953	0.3589	0.0119
KNN	0.943	0.943	0.945	0.943	0.3316	0.0586
NB	0.930	0.930	0.935	0.930	0.4164	0.0312
SVM	0.590	0.590	0.682	0.543	0.8438	0.0953

Here, to minimize the angle-dependent scattering effects caused by particle morphology in both the single and multiple scattering regimes, an unpolarized light source was employed, and all scattered light (i.e., the scattering matrix element S11) was measured regardless of the polarization direction.

ASDI, which drives a 1D LED array and PDs instead of a detector array^56,57 or an image sensor^58,59 for MALS measurement, features a compact design and cost-effective manufacturing due to simplified components. In addition, excellent classification performance has been verified with commonly used machine learning algorithms. In particular, the DT classification algorithm achieved a high accuracy of 95% in identifying five types of fire and five types of non-fire aerosols. The DT classification algorithm with simplified nodes enables multi-class classification at the board level for on-device implementation, which is highly favorable for practicality and product commercialization. The structure of the DT algorithm is inserted in Fig. S4.

In order to enhance classification accuracy, accurate measurement of optical properties is indispensable, which requires multi-angle LED irradiation over a wide range. The maximum illumination angle is determined by LED position, LED viewing angle, and the numerical aperture of the lens. The effective illumination angle range of ASDI’s sensor module is approximately 28 degrees, and operating multiple LEDs in a 2D array within the effective illumination further improves accuracy and sensitivity of scattering coefficient estimation and enhance classification performance, but requires reasonable trade-offs in terms of cost, power consumption, and size. By adding light sources with different wavelengths, such as ultraviolet and infrared, a greater number of measurement variables for the optical property ratios, which serve as the fingerprints of aerosols, can be obtained to recognize more colorful fire and non-fire aerosols.

In addition to reducing false alarms, this study is expected to identify five classes of fires–ordinary fires, burning liquid or gas, electrical fires, metallic fires, and grease or cooking fires–and suggest appropriate fire extinguishing methods for each fire class. The feasibility of reducing false alarms was primarily evaluated by focusing on fire smoke generated from single materials and non-fire aerosols that frequently cause nuisance alarms. The experiments were performed under conditions where the smoke concentration was below 20%/m, corresponding to an optical thickness of less than approximately 0.22, where single scattering is more dominant than multiple scattering. This study is expected to serve as a foundation for future research on smoke detection systems capable of accurately distinguishing between fire and non-fire aerosols in complex fire scenarios, such as the simultaneous combustion of multiple materials or the coexistence of fire and non-fire aerosols, where multiple scattering becomes dominant.

Methods

LED array-based sensor module for ASDI

The center wavelength (CW), full width at half maximum (FWHM), and viewing angle (VA) of the near-infrared LEDs (Vishay Intertechnology, VSMY5940X01) –LED1, LED2, and LED3–are 940 nm, 50 nm and , respectively. LED4, which has a CW of 542 nm, FWHM of 197 nm and VA of , is a white LED (Bivar, SME2014UWDN05), and LED5 and LED6 are blue LEDs (SunLED, XZCB25X109FS) with a CW of 450 nm, FWHM of 15 nm and VA of . The emission spectra of all LEDs are visualized in Fig. S5. The focal length and the diameter of the lens is 15 mm and 20 mm, respectively. PD1 and PD2 are the same photodiode (ams OSRAM, SFH 2200). The aperture size is 4 mm. ASDI’s sensor module case was custom-made using a 3D printer. The LED array board, photodiode module, and board for control and signal acquisition were custom designed and manufactured. To enhance the sensitivity of LED light signal measurements, the light intensity of each LED was individually controlled, and the PD gain was set differently for each LED. Specifically, the light intensities from LED1, 2 and LED5, 6 were maximized and detected with high gain, while the light intensities from LED3 and LED4 were adjusted to avoid saturation and measured with low gain. All analog signals were converted into 21-bit digital signals and recorded.

Sample preparation for optical property measurement

Each particle employed for sample preparation was mixed in PDMS-A (Dow Corning, Silicone elastomer base) with a specific weight ratio and then stirred for 3 hours using a magnetic stirrer to obtain a uniform dispersion. PDMS-B (Dow Corning, Silicone elastomer curing agent) was added to the particle-mixed solution in a 10:1 ratio to PDMS-A and then stirred for 10 minutes. The degassed solution was poured into a petri dish and naturally cured at room temperature for 48 hours.

Samples with various particles were fabricated using PS (Sigma-Aldrich, 89904), PMMA (Sigma-Aldrich, 90875), and carbon (Sigma-Aldrich, 633100) particles. To fabricate samples with different particle sizes, PS particles of 100 nm (Sigma-Aldrich, 43302), 200 nm (Sigma-Aldrich, 69057), 500 nm (Sigma-Aldrich, 59769), and 1 m (Sigma-Aldrich, 89904) were used.

Experimental setup of ASDI for aspirating smoke detection

Experiments were performed on five types of fire sources (filter paper, kerosene, PVC, ABS, cotton wick) and five types of non-fire aerosols (water vapor, oil mist, smoke check spray, fog machine, fly ash) at a temperature of 24 C and a relative humidity of 30%. The airflow velocity in the pipe was approximately 3 m/s. All materials, except for fly ash, were inspected in the identical chamber, while fly ash was tested in a separately designed custom chamber.

A filter paper (ADVANTEC, 00021090) was placed on a hot plate heated to 400 C, and a presser was applied on top to prevent it from lifting during the smoking process. In the kerosene experiment, smoke was generated by igniting a wick-tape alcohol lamp filled with kerosene, and the flame was blocked using a ceramic wire mesh. A PVC sheet (SUNKYUNG PLATECH, ivory PVC) and an ABS sheet (SUNKYUNG PLATECH, white ABS) were directly ignited with a gas torch to generate smoke, and both sheets had a thickness of 2 mm. Cotton wicks, cut into 5 mm lengths, were directly burned with a gas torch to generate smoke. Water vapor was examined by boiling water in an electric kettle. Oil (CJ CheilJedang, Extra-virgin olive oil) was poured into a frying pan and heated outside the chamber using a portable butane gas stove. Once smoke was generated, the frying pan was moved inside to observe the oil mist. A smoke check spray (HSI Fire & Safety Group, SmokeCheck) and a fog machine (beamZ, S1500) were manually sprayed for 5 seconds and 2 seconds, respectively. In fly ash (Jinsol Dust, Class 5) experiment, a separate chamber was custom-built to float the fly ash and circulate it within the chamber and pipe (Fig. S6). Fly ash was placed on weighing paper under a fan located at the center of the chamber. Fly ash was detected by first driving a fan that sucked the fly ash into a pipe, and then operating another fan that floated the fly ash into the air.

Acknowledgements

This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2024-00461079, Development of Adaptive On-Device Software Technology for Environmental Adaptation in Unmanned Vehicle Surveillance Equipment).

Author contributions

S.K. conceived the idea for the project and developed it with support from all other authors. S.K. and H.Y. designed the smoke detection instrument and experimental environments for fire and non-fire smoke experiments. H.Y. designed the printed circuit board for the sensor module. K.C. and K.H. developed the software for operating the sensor module. S.K., K.H., J.H.R., and Y.A. conducted the investigations and experiments. S.K. performed signal analysis and created the visualizations. H.Y. and K.L. acquired funding. S.K. wrote the original draft. All authors contributed to the review and editing of the manuscript.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-025-11185-6.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. Fleming, J. M. Photoelectric and ionization detectors-A review of the literature re-visited. Retrieved31, 2010 (2004).

2. Cleary, T. Performance of dual photoelectric/ionization smoke alarms in full-scale fire tests. Fire Technol.; 2014; 50, pp. 753-773.

3. Wei, M.-C., Lin, B.-R., Lin, Y.-Y., Chiou, G.-J. & Kuo, W.-K. Experimental study on effects of light source and different smoke characteristics on signal intensity of photoelectric smoke detectors. In 2021 IEEE 3rd Eurasia Conference on IOT, Communication and Engineering (ECICE). 518–522 (IEEE, 2021).

4. Tannous, WK. The economic cost of unwanted automatic fire alarms. Fire Saf. J.; 2021; 124, 103394.

5. Ahrens, M. False alarms and unwanted activations” from us experience with smoke alarms and other fire detection/alarm equipment. (National Fire Protection Association, 2004).

6. Li, K et al. Dual-wavelength smoke detector measuring both light scattering and extinction to reduce false alarms. Fire; 2023; 6, 140.

7. Han, K; Kim, S; Yang, H; Cho, K; Lee, K. Time series classification with multiple wavelength scattering signals for nuisance alarm mitigation. Fire; 2023; 7, 14.

8. Uthe, EE. Particle size evaluations using multiwavelength extinction measurements. Appl. Opt.; 1982; 21, pp. 454-459.1982ApOpt.21.454U1:STN:280:DC%2BC3c3ktlaqsA%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/20372477]

9. Cashdollar, KL; Lee, CK; Singer, JM. Three-wavelength light transmission technique to measure smoke particle size and concentration. Appl. Opt.; 1979; 18, pp. 1763-1769.1979ApOpt.18.1763C1:STN:280:DC%2BC3c7mslSnsg%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/20212546]

10. Loepfe, M; Ryser, P; Tompkin, C; Wieser, D. Optical properties of fire and non-fire aerosols. Fire Saf. J.; 1997; 29, pp. 185-194.

11. Wen-qing, W et al. Multi-angle scattering characteristic of test fire smoke and typical interference aerosol. Proc. Eng.; 2011; 11, pp. 466-472.

12. Weinert, DW; Cleary, TG; Mulholland, GW; Beever, PF. Light scattering characteristics and size distribution of smoke and nuisance aerosols. Fire Saf. Sci.; 2003; 7, pp. 209-220.

13. Meacham, BJ; Motevalli, V. Characterization of smoke from smoldering combustion for the evaluation of light scattering type smoke detector response. J. Fire Protect. Eng.; 1992; 4, pp. 17-32.

14. Schultze, T., Sichma, L. & Meyer, M. A smoke detector to prevent false alarms in lunar missions by smoke-dust discrimination. In Proceedings of the 2020 International Conference on Environmental Systems (2020 International Conference on Environmental Systems, 2020).

15. Cleary, T. & Mensch, A. Polarized light scattering of smoke sources and cooking aerosols. In AUBE’17 (2017).

16. Krüll, W. 16th International Conference on Automatic Fire Detection & Suppression, Detection and Signaling Research and Applications Conference: Proceedings: September 12–14, 2017, College Park Marriott Hotel & Conference Center, Hyattsville, MD, USA (University Duisburg-Essen, Communication Systems, NTS, 2017).

17. Marcius, L., Schultze, T. & Willms, I. Design of a polarimetric scattering-based aerosol-classifying smoke detector prototype. AUBE17 (2017).

18. Gottuk, DT; Peatross, MJ; Roby, RJ; Beyler, CL. Advanced fire detection using multi-signature alarm algorithms. Fire Saf. J.; 2002; 37, pp. 381-394.1:CAS:528:DC%2BD38XitlWhu74%3D

19. Chen, S-J; Hovde, DC; Peterson, KA; Marshall, AW. Fire detection using smoke and gas sensors. Fire Saf. J.; 2007; 42, pp. 507-515.1:CAS:528:DC%2BD2sXht1eqsL3P

20. Chen, X. & Bu, L. Research of fire detection method based on multi-sensor data fusion. In 2010 International Conference on Computational Intelligence and Software Engineering. 1–4 (IEEE, 2010).

21. Wang, S; Xiao, X; Deng, T; Chen, A; Zhu, M. A Sauter mean diameter sensor for fire smoke detection. Sens. Actuators B Chem.; 2019; 281, pp. 920-932.2019SeAcB.281.920W1:CAS:528:DC%2BC1cXit1ehurzI

22. Kheirkhah, P; Baldelli, A; Kirchen, P; Rogak, S. Development and validation of a multi-angle light scattering method for fast engine soot mass and size measurements. Aerosol Sci. Technol.; 2020; 54, pp. 1083-1101.2020AerST.54.1083K1:CAS:528:DC%2BB3cXhtVWmtLrO

23. Link, O; Snelling, D; Thomson, K; Smallwood, G. Development of absolute intensity multi-angle light scattering for the determination of polydisperse soot aggregate properties. Proc. Combus. Inst.; 2011; 33, pp. 847-854.1:CAS:528:DC%2BC3MXptFGhsLk%3D

24. Chen, D et al. A new angular light scattering measurement of particulate matter mass concentration for homogeneous spherical particles. Sensors; 2019; 19, 2243.2019Senso.19.2243C1:CAS:528:DC%2BB3cXpvVeksQ%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31096589][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6567324]

25. Pittman, ZA; McCarthy, ME; Birtwistle, MR; Kitchens, CL. Method for improved fluorescence corrections for molar mass characterization by multiangle light scattering. Biomacromolecules; 2022; 23, pp. 3743-3751.1:CAS:528:DC%2BB38XitVagsbfK [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35926160][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9843603]

26. Keller, A; Loepfe, M; Nebiker, P; Pleisch, R; Burtscher, H. On-line determination of the optical properties of particles produced by test fires. Fire Saf. J.; 2006; 41, pp. 266-273.1:CAS:528:DC%2BD28XkslKrtr0%3D

27. Shao, W; Zhang, H; Zhou, H. Fine particle sensor based on multi-angle light scattering and data fusion. Sensors; 2017; 17, 1033.2017Senso.17.1033S [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28471406][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5469638]

28. Yu, X; Shi, Y; Wang, T; Sun, X. Dust-concentration measurement based on Mie scattering of a laser beam. PLoS One; 2017; 12, e0181575. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28767662][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5540421]

29. Lin, M; Zhu, M; Xiao, X; Li, C; Wu, J. Optical sensor for combustion aerosol particle size distribution measurement based on embedded chip with low-complexity mie scattering algorithm. Opt. Laser Technol.; 2023; 158, 108791.1:CAS:528:DC%2BB38XivVSit77I

30. Lin, M. et al. Optical scattering sensing for combustion smoke aerosol shape and particle size distribution using classification-regression concatenated framework network. Available at SSRN 5233420 .

31. Li, S et al. Multiple scattering of light transmission in a smoke layer. Optik; 2014; 125, pp. 2185-2190.2014Optik.125.2185L1:CAS:528:DC%2BC2cXhtFOnu7g%3D

32. Liu, F; Zhao, Z; Yao, H-W; Liang, D. Application of aspirating smoke detectors at the fire earliest stage. Proc. Eng.; 2013; 52, pp. 671-675.

33. Johnson, P; Beyler, C; Croce, P; Dubay, C; McNamee, M. Very early smoke detection apparatus (Vesda), David Packham, John Petersen, Martin Cole: 2017 Dinenno prize. Fire Sci. Rev.; 2017; 6, 5.

34. Reynolds, K; De Kock, J; Tarassenko, L; Moyle, J. Temperature dependence of led and its theoretical effect on pulse oximetry. Br. J. Anaesth.; 1991; 67, pp. 638-643.1:STN:280:DyaK38%2FovVyjsQ%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/1751282]

35. Rada, NM; Triplett, GE. Thermal and spectral analysis of self-heating effects in high-power leds. Solid-State Electron.; 2010; 54, pp. 378-381.2010SSEle.54.378R1:CAS:528:DC%2BC3cXivVyrt7g%3D

36. Fabian, T. Z. & Pravinray, D. G. Smoke Characterization Project (Fire Protection Research Foundation/Underwriters Laboratories, Incorporated, 2007).

37. Hohmann, M et al. Direct measurement of the scattering coefficient. Biomed. Opt. Exp.; 2020; 12, pp. 320-335.

38. Deng, T; Wang, S; Xiao, X; Zhu, M. Eliminating the effects of refractive indices for both white smokes and black smokes in optical fire detector. Sens. Actuators B Chem.; 2017; 253, pp. 187-195.2017SeAcB.253.187D1:CAS:528:DC%2BC2sXhtVGrsr%2FE

39. Dong, W-H; Sheng, X-E; Wang, S; Deng, T. Experimental study on particle size distribution characteristics of aerosol for fire detection. Appl. Sci.; 2023; 13, 5592.1:CAS:528:DC%2BB3sXhtVSrs7vN

40. Teri, M et al. Impact of particle size, refractive index, and shape on the determination of the particle scattering coefficient-an optical closure study evaluating different nephelometer angular truncation and illumination corrections. Atmos. Meas. Tech.; 2022; 15, pp. 3161-3187.1:CAS:528:DC%2BB38XitlaqtrnE

41. Jones, A. Light scattering for particle characterization. Prog. Exp. Study Part. Size Distrib. Character. Aerosol Fire Detect. Energy Combus. Sci.; 1999; 25, pp. 1-53.

42. Khlebtsov, N; Trachuk, L; Mel’nikov, A. The effect of the size, shape, and structure of metal nanoparticles on the dependence of their optical properties on the refractive index of a disperse medium. Opt. Spectrosc.; 2005; 98, pp. 77-83.2005OptSp.98..77K1:CAS:528:DC%2BD2MXhslakt74%3D

43. Piederrière, Y et al. Scattering through fluids: Speckle size measurement and Monte Carlo simulations close to and into the multiple scattering. Opt. Exp.; 2004; 12, pp. 176-188.

44. Unosson, J. Physical Properties of Acidic Calcium Phosphate Cements. Ph.D. Thesis. (Acta Universitatis Upsaliensis, 2014).

45. Kennard, E. H. et al. Kinetic Theory of Gases. Vol. 483 (McGraw-Hill, 1938).

46. Ma, L et al. Dependent scattering and absorption by densely packed discrete spherical particles: Effects of complex refractive index. J. Quant. Spectrosc. Radiat. Transf.; 2017; 196, pp. 94-102.2017JQSRT.196..94M1:CAS:528:DC%2BC2sXlvVGgt70%3D

47. Farcas, MT et al. Evaluation of pulmonary effects of 3-D printer emissions from acrylonitrile butadiene styrene using an air-liquid interface model of primary normal human-derived bronchial epithelial cells. Int. J. Toxicol.; 2022; 41, pp. 312-328.1:CAS:528:DC%2BB38XhvVCgs7jK [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35586871]

48. Tsai, L-H; Yang, PN; Wu, C-C; Lin, HY. Quantifying scattering coefficient for multiple scattering effect by combining optical coherence tomography with finite-difference time-domain simulation method. J. Biomed. Opt.; 2018; 23, pp. 086004-086004.2018JBO..23h6004T

49. Sorensen, C. Light scattering by fractal aggregates: A review. Aerosol Sci. Technol.; 2001; 35, pp. 648-687.2001AerST.35.648S1:CAS:528:DC%2BD3MXlvVSkurc%3D

50. Fujii, H; Na, H; Yi, J; Kobayashi, K; Watanabe, M. Particle size distribution effects on the light scattering properties in non-diluted colloidal suspensions: A numerical study. Colloids and Surf. A Physicochem. Eng. Asp.; 2024; 703, 135208.1:CAS:528:DC%2BB2cXhvVymsbnJ

51. Mayoraz, E. & Alpaydin, E. Support vector machines for multi-class classification. In International Work-Conference on Artificial Neural Networks. 833–842 (Springer, 1999).

52. Hsu, C-W; Lin, C-J. A comparison of methods for multiclass support vector machines. IEEE Trans. Neural Netw.; 2002; 13, pp. 415-425. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/18244442]

53. Nousiainen, T; Zubko, E; Lindqvist, H; Kahnert, M; Tyynelä, J. Comparison of scattering by different nonspherical, wavelength-scale particles. J. Quant. Spectrosc. Radiat. Transf.; 2012; 113, pp. 2391-2405.2012JQSRT.113.2391N1:CAS:528:DC%2BC38XlvFKgt70%3D

54. Mishchenko, MI; Hovenier, JW; Travis, LD. Light scattering by nonspherical particles: Theory, measurements, and applications. Meas. Sci. Technol.; 2000; 11, pp. 1827-1827.2000MeScT.11.1827M

55. Kahnert, M. Electromagnetic scattering by nonspherical particles: Recent advances. J. Quant. Spectrosc. Radiat. Transf.; 2010; 111, pp. 1788-1790.2010JQSRT.111.1788K1:CAS:528:DC%2BC3cXnt1yntbw%3D

56. Bartholdi, M; Salzman, G; Hiebert, R; Kerker, M. Differential light scattering photometer for rapid analysis of single particles in flow. Appl. Opt.; 1980; 19, pp. 1573-1581.1980ApOpt.19.1573B1:STN:280:DC%2BC3c7ns12jsg%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/20221079]

57. Dick, WD; Ziemann, PJ; McMurry, PH. Multiangle light-scattering measurements of refractive index of submicron atmospheric particles. Aerosol Sci. Technol.; 2007; 41, pp. 549-569.2007AerST.41.549D1:CAS:528:DC%2BD2sXltlalt7Y%3D

58. Chen, A et al. Light scattering intensity field imaging sensor for in situ aerosol analysis. ACS Sens.; 2020; 5, pp. 2061-2066.1:CAS:528:DC%2BB3cXhtlSisLbP [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32608223]

59. Oltmann, H; Reimann, J; Will, S. Wide-angle light scattering (wals) for soot aggregate characterization. Combust. Flame; 2010; 157, pp. 516-522.2010CoFl.157.516O1:CAS:528:DC%2BC3cXmvFCmtA%3D%3D