A NDIR sensor is a simple sensor that can analyze the spectral information of a sample without dispersive elements such as a diffraction grating or prism [1,2]. Owing to its fast response speed, high accuracy, and low cost, it is utilized in various applications such as fire detection [3,4], air quality monitoring [5], and breathing analysis [6]. Particularly in fire detection, researchers have reported studies that combine NDIR sensors with multiple sensors [7], or use a fusion algorithm [8] to reduce nuisance alarms and the response time to fires, distinguish between fires and non-fires, and enhance detection accuracy.

Although many studies for improving the performance of fire detectors have been actively reported [9,10], studies on fire detectors that detect types of fire are insufficient. Detection of fire types is meaningful for rapidly extinguishing a fire and preventing the spread of fire. The fire safety standards for fire extinguishers and automatic fire extinguishing systems (NFSC 101) classify fires into Class A- general fires, Class B- oil fires, Class C- electrical fires, Class D- metal fires, and Class K- kitchen fires, and suggest specific extinguishing methods suitable for each type of fire. For example, foam and powder or an electrical insulating extinguishing agent that can block oxygen is required to extinguish oil and electrical fires. Using an unsuitable extinguishing agent such as water in Class B and C fires will accelerate its spread and cause secondary injuries from electric shock. Accordingly, detecting the type of fire is essential to swiftly extinguish the fire and minimize fire damage. Here, we present a method to detect and classify aerosol particles and gases using the NDIR-based dual-channel mid-IR sensor. To validate this, the spectral transmittances (T_λ) of mid-IR LEDs for aerosol particles and gases were measured and T_λ and a ratio between spectral transmittances (Ratio_T) were used to classify aerosol particles and gases via hierarchical clustering. Actively applying this technique to fire detectors is expected to enable a rapid detection of the fire type by analyzing fire smoke, thus helping to quickly extinguish fires and prevent their spread.

To verify the feasibility of the detection and classification of aerosol particles and gases using the NDIR-based dual-channel mid-IR measurement, we performed experiments measuring the transmittance of various aerosol particles and gases, and then classified the aerosol particles and gases using hierarchical cluster analysis. Figure 3 presents the measured transmittance (i.e., T_{4.2 μm} and T_{4.7 μm}) results for the samples. The solid blue and orange lines indicate the measured T_{4.2 μm} and T_{4.7 μm}, respectively, and the black arrows indicate the starting point of the sample inflow. The transmittances for 4.2 μm light and 4.7 μm light were observed to approximately 1 at the absence of sample inflow (Figure 3(a)). With gas inflow when the lighter flame is ignited, T_{4.7 μm} is constant at approximately 1, whereas is T_{4.2 μm}slightly decreased (Figure 3(b)). The 4.2 μm mid-IR LED light source with a wide bandwidth also emits 3.4 μm light [11], which was absorbed by the butane gas [12]. The measured transmittances of the smoke detector tester spray which consists of isobutane, propane, siloxane, and butane gases are exhibited in Figure 3(c). Both T_{4.2 μm} and T_{4.7 μm} decreased owing to the multiple gases. For CO₂ gas, which has a strong absorption spectrum at approximately 4.26 μm, T_{4.2 μm}decreased substantially, and a slight decrease was observed at T_{4.7 μm} because of the wide bandwidth of the 4.7 μm LED light source (Figure 3(d)). The absorption spectrum of water highly depends on its molecular state (solid, liquid, or gas) [13,14]. For water vapor, both T_{4.2 μm} and T_{4.7 μm} were significantly dropped due to strong IR absorption (Figure 3(e)).

Figure 4(a) illustrates a three-dimensional (3D) graph of the values of T4.2 μm, T_{4.7 μm}, and Ratio_T measured in various samples. T_{4.2 μm} and T_{4.7 μm} were calculated from minimum values measured when the sample was sprayed. A hierarchical cluster analysis was performed to classify the samples by calculating the distance between the clusters. The distance can be expressed as

where d is the distance between clusters, r and s are objects, x¯r and x¯s are the cluster centers, and n_r and n_s are the numbers of elements. Air (i.e., free space), water vapor, and CO₂ gas clusters were separated from each other, whereas for the lighter and the smoke detector tester spray, a few objects overlapped with each other. This is likely because they share a similar gas composition with butane.

We detected and classified aerosol particles and gases employing the NDIR-based dual-channel mid-IR measurement. 4.2 μm and 4.7 μm mid-IR LEDs were used for light sources, which were modulated with 900 Hz and 1,000 Hz, respectively to reduce interference between the light sources and ambient light, and increase SNR. The modulated 4.2 μm and 4.7 μm lights passed through lenses and were acquired by a photodetector, and the transmittances (T_{4.2 μm}, T_{4.7 μm}) were obtained to detect the aerosol particles and gases. The aerosol particles and gases were classified by the distance between clusters of objects via T4.2 μm, T_{4.7 μm} and Ratio_T. To detect CO and CO₂, which are representative gases generated during hydrocarbon combustion, we utilized mid-IR LED light sources in the 4.2 μm and 4.7 μm bands, which can be strongly absorbed by the gases. In an actual fire, a variety of gases are produced depending on the combustible’s chemical properties and combustion state. For example, carbon monoxide, carbon dioxide, hydrogen cyanide, and ammonia are generated in fires caused by general heating materials such as wood and textiles (Class A- general fires); propane gas (Class B- oil fires); hydrogen fluoride, hydrogen chloride, sulfur dioxide, and carbon monoxide in lithium-ion battery fires (Class C- electrical fires); and hydrogen gas (Class D- metal fires). Therefore, to detect the fire type, it is critical to select a light source suitable for gas detection. The acquisition speed is determined by the sampling rate of the photodetector (up to 9 MHz) and the number of data acquired. In this study, data were acquired at a rate of 500 kHz and the aerosol particles and gases were analyzed at 0.25-second intervals using a window size of 250,000 samples and an overlap of 125,000 samples. Propane gas, CO₂ gas, and water vapor were classified through a hierarchical cluster analysis of the data expressed as T4.2 μm, T_{4.7 μm}, and Ratio_T values. By measuring multiple spectral absorbances, analyzing the relationship between absorbances, and applying a machine learning or deep learning-based classification algorithm, it is expected to accurately classify various types of aerosol particles and gases. This can be applied in fire detection to develop a highly useful technique that can classify fire types and distinguish between fires and non-fires. Furthermore, the NDIR-based dual-channel mid-IR method is low-cost and simple and can perform high-speed real-time measurements. Therefore, in addition to fire type detection, it is expected to be actively employed in many applications such as monitoring toxic gases generated in industrial sites.