Introduction

Forest fires not only directly destroy forest resources and often cause a large number of casualties, but also indirectly deteriorate the global ecological environment and exacerbate the greenhouse effect. Some countries in the Americas, Australia, Africa, and Asia are areas where forest fires occur frequently. Among many influencing factors, deficiencies in firefighting equipment has become a main reason for difficulties of forest fire fighting in several countries. At present, the most advanced space-air-ground integrated fire-fighting system plays an important role in monitoring and extinguishing forest fires and can slow down the consumption of forest resources to some extent. But the cost of the equipment is considerable. Therefore, firefighting with aircraft has become an important choice for combatting forest fires in many countries.

In forest fire management, residual clean-up is essential. Preventing secondary burning caused by the re-ignition of residual fires not only saves fire-fighting time but also reduces fire casualties. Residual fires in forests includes visible flames on trees and smoldering wood on the ground. Fire and exposed smoldering wood are easy to find and clean, but hidden smoldering wood under the ground is often difficult to detect. For example, smoldering wood is covered by ash, leaves, soil, smoldering roots, and other debris. In order to facilitate the later description, the former is α (exposed smoldering) and the latter is β (hidden smoldering). As shown in Table 1, a β-fire can also develop downward into peat fire under suitable humus moisture content, increasing the duration of underground fire (Han et al. 2022).

Table 1 Classification of smoldering fire

There has been considerable research on residual fire detection. From the perspective of thermal radiation, He et al. (2018) determined residual fires based on the standard deviation of the temperature data of an infrared image. The method is straight-forward and efficient with significant advantages, but only suitable for rough filtering of suspected smoldering points. This is because the change of temperature in the forest environment is not just affected by heat radiation of the residual fire. Factors such as the amount of sunlight, thermostatic animals, ground heat and ground cover can all contribute to data interference. Levoglucosan is a typical product of smothering of cellulose-based combustibles and it can be used as the detection index of smoldering objects (Madsen et al. 2018). This study provides a basis for the detection of cellulose smoldering but is not suitable for the detection of forest fires. A drone equipped with a high-definition camera, or an infrared camera identifies smoke by image processing to find the fire point (Fan and Ma 2015; Zheng and Zhai 2015; Gao and Cheng 2019; Sun et al. 2021; Hu et al. 2022; Kwak and Ryu 2022; Li et al. 2022). In practice, a piece of smoldering charcoal does not produce significant temperature changes and smoke characteristics on the surface above it but can still be transformed into an open flame under certain conditions. Figure 1 illustrates the conversion of a β-fire to an open flame, showing the characteristics of weak smoldering and long duration.

Fig. 1
figure 1

Reignition of a β-fire: a charcoal for overcast burning; b smoke appears 12 min after adding leaf litter; c no significant increase in surface temperature 15 min after adding leaf litter; and d dead leaves covering the embers burn in 1 h

A forest fire detection robot is the main research direction of forest fire cleaning equipment. The related research focuses on the design of the cleaning robot (Zhang and Jiang 2015; Yao 2021; Yang et al. 2022), actuator control (Xu and Jiang 2017; Zhu 2017), multi-sensor detection system (Wang et al. 2010; Song 2017) and robot path planning (Liu 2019; Liu et al. 2020). A fire detection robot can enter a dangerous environment, replacing people, but existing detection methods make it better for detecting of α-type fires and insufficient for β-type fires. Smoldering charcoal produces more than twice as much CO gas as normal combustion (Zhao 2014), its concentration is an important detection index of fire detection (Jiang et al. 2016; Qu et al. 2017; Dang et al. 2018; Fonollosa et al. 2018; Conceição et al. 2020). Tang et al. (2022) used variance analysis to note that smoldering depth has a significant effect on nitrogen oxide emissions but not on CO emissions. CO gas molecules easily pass through soil, wood ash, dead leaves, and other ground covers and may be suitable as β-type fire detection indicators. Moreover, compared with levoglucosan, CO detection equipment has advantages of low price, small size and reliability. Therefore, it is of significance to study the CO gas concentration in the air near the ground above a β-type residual fire by using CO sensors, and to design a fire detection device based on CO gas detection to locate β-type residual fires.

Materials and methods

Materials and CO sensor

Ezo spruce [Picea jezoensis (Siebold & Zucc.) Carr.] is a common, low resin shallow rooted species in northern forests suitable for this study. Wood pieces with 80-mm diameter (from Tanglin Forest Farm) were sawn into 200 mm segments and burnt completely one by one in a 400 × 400 pit, 500 mm deep (Fig. 1) to produce smoldering wood carbon (Fig. 2).

Fig. 2
figure 2

Materials for smoldering: a Picea jezoensis piece; b experimental length; c diameter of experimental material; and d preparation of smoldering wood carbon

This study used a voltage-type CO sensor to output an analog signal which was converted into a digital signal. The CO sensor was connected to the development board, and the data transmission between the CO sensor and the computer recognized through a Bluetooth module. CO sensors use electrochemical probes with higher accuracy and stability than conventional semiconductor probes. The main parameters are shown in Table 2.

Table 2 Main parameters of CO sensor

Experimental principles

Gas diffusion

CO density is less than the atmosphere and in a windless state, will freely diffuse upward after passing through the soil layer, showing a ‘V’ shape. The diffusion law of CO gas below 1200 mm above ground is affected by the installation height of the detection device robot and is the focus of this study. The six carbon monoxide sensors were evenly arranged above the smoldering charcoal at 200 mm apart. The concentration of CO at different heights was measured vertically and the relationship was determined and used as a basis for identifying an underground smoldering fire.

Motion decomposition search algorithm

Moving the detection device to a position directly above the suspected point is based on judgment. The detection device was set to search step by step in a traverse direction when CO gas was not detected (Fig. 3a). In a windless state, when the detection device satisfies the relationship between CO concentration and height, it can be confirmed that there is smoldering fire on the ground directly below the detection device (Fig. 3b).

Fig. 3
figure 3

Detection device search diagram under windless conditions: a traverse search; b device detects residual fire

Under the influence of a one-way weak wind, the rising path of CO was θ angle (< 90°) with the ground. When the device detected CO and stopped traversing, the position directly below was still some distance from the β-type residual fire. The key to solving the problem is to control the detection device to reach a position directly above the β-type residual fire along the CO rising pathway. As shown in Fig. 4, according to the requirements of the size and spacing distance of the CO sensor, the detection device is designed as a square. Four CO sensors were arranged on the front of the detection device (Fig. 4a), and the other four arranged inside the device at intervals of 100 mm (Fig. 4b).

Fig. 4
figure 4

Detection device: a overall appearance; b internal structure

The diffusion of CO from the smoldering charcoal above the ground has two characteristics. One is that the concentration of CO is highest in the center of the gas mass on a horizontal surface, and the concentration of carbon monoxide around it gradually decreases with increasing distance. The second is that the concentration of CO is highest at the ground and the concentration of CO at other points decreases with distance from the ground level. Based on the above characteristics, a reverse search algorithm is designed to enable the detection device to find the smoldering charcoal as it gradually approaches the highest point of carbon monoxide concentration. It is stipulated that the detection device moves the same step h each time. Directly compare the measured values of the four CO sensors on the front of the detection device and use the following rules to control the movement of the detection device. Comparing the values of sensors 1 and 3 in the horizontal direction, the detection device moves in the direction of the largest values until the conditions are not satisfied. Comparing the values of sensors 2 and 4 in the vertical direction, the detection device moves in the direction of large values until the conditions are not satisfied. Following these steps, the measurements of sensors 5, 6, 7, and 8 are verified to meet the laws of CO concentration and height. The above three steps are repeated until the device detects the smoldering fire or not.

Experiment setup

Introduction of experimental environment

To avoid the influence of wind, all experiments were carried out in a square, 1250 mm × 800 mm × 1200 mm topless combustion chamber (Fig. 5) composed of glass windows. An aluminum combustion basin (120 mm high) was placed in the center of the bottom of the combustion chamber to collection smoldering wood carbon and the mixture of soil and ash. During the experiment, to avoid the local heat cracking of the bottom glass, the items in the combustion basin were placed in the order of 30 mm thick dry soil, freshly prepared smoldering wood carbon, and the mixture of ash and soil (1:1). A fan with adjustable height was installed on the column of the smoldering chamber to provide one-way wind.

Fig. 5
figure 5

Experimental environment

Apparatus for studying the relationship between CO concentration and height

Detection device I (Fig. 6) was used for studying the variation of CO concentration with height. Copper pillars (M3) connected six CO sensors in series so that the distance between adjacent sensors was 200 mm (8 copper pillars). Two screws fixed the connected CO sensor columns in the middle of a 1250 mm × 50 mm × 20 mm Board. The two ends of the board were placed in the middle of the upper end of the two opposite faces in the long direction of the smoldering chamber so that the CO sensor column was located above the center of the combustion basin. The development board (MEGA2560 R3) connected to six carbon monoxide sensors transmitted the data to the computer via Bluetooth. A mobile battery powered the sensor and development board.

Fig. 6
figure 6

Detection device I

Experimental apparatus for testing motion decomposition algorithm

The verification device of the motion decomposition algorithm is composed of detection device II, mobile device, and smoldering chamber (Fig. 4). The end of the CO sensor series was connected to the support end of the 3D printed part by copper pillars. The fixing end of the 3D printed part was connected to the inside of the ABS housing backplane by two bolts to ensure that the four CO sensors were located exactly at the 60, 160, 260 and 360 mm axes inside the housing. The mobile device was composed of guide rail, slider, motor, pan-tilt, and battery. The detection device II was connected to the pan-tilt through a connector. Under the control of the mobile device, detection device II has the function of moving in front and back, left and right, up and down and rotates in the horizontal direction. Two ends of the X-direction guide rails of the mobile device were connected to the top of two opposite surfaces in the longitudinal direction of the smoldering chamber through a support (Fig. 7). Figure 8 shows the experimental setup.

Fig. 7
figure 7

Detection device II and mobile device

Fig. 8
figure 8

Motion decomposition algorithm verification device

Results

Characteristics of CO concentration above the β-type residual fire

Default smoldering was stable and smoldering characteristics regular. To reduce the influence of experimental environment on the smoldering process, only data in the first 105 s were taken. The results of each CO sensor in detection device I within 105 s are plotted in Fig. 9.

Fig. 9
figure 9

Carbon monoxide concentration at each measuring point

The CO concentration at each measurement point constantly changed with time and height of 200 mm sensor measurement of the maximum value of 17.8 × 10−3 mL m−3 and the minimum value of 2.2 × 10−3 mL m−3 are shown in Fig. 9. With the increase in height, CO concentration curve gradually flattened. The six curves in region R were clearly layered and in the same period of time, CO concentrations at different heights were significantly different. Further analysis of the data between average CO concentration and height of each measurement point is shown in Fig. 10.

Fig. 10
figure 10

Relationship between average carbon monoxide concentration and height at each measuring point: a linear fitting of 10–15 s data; b linear fitting of 20–25 s data; c linear fitting of 0–105 s data; d exponential fitting of 10–15 s data; e exponential fitting of 20–25 s data; and f exponential fitting of 0–105 s data

The scatter plot of the average value of each sensor measurement with height change in the selected time period was drawn and linear and exponential fitting analysis were performed in each time period. From the value of R2, it can be seen that the results of exponential fitting were better than those of linear fitting in both the long time (105 s) and short time periods (5 s). Therefore, in a windless state, the mean value of CO concentration in the near-surface space (1.2 m) of a β-type residual fire decreases exponentially with the increase of height.

Search algorithm

Combined with the above results, the law that the mean value of CO concentration decreases exponentially with height was used as the basis of the motion decomposition algorithm. Programs based on the Arduino platform were written and five groups of experiments carried out with search steps of 1, 3, 5, 10 and 20 cm. Each group was repeated five times. The results are shown in Table 3.

Table 3 Experimental results for five groups with different search steps

The results show that the search step size should be controlled within a reasonable range; too large or too small will result in detection failure. When the search step was 1 cm, detection device II eventually stopped in the search path and did not reach the top of the combustion basin. When the search step was 3 or 5 cm, the detection device II successfully detected the smoldering fire. In the eight test failures of group 4 and 5, detection device II finally showed the characteristics of crossing the combustion basin and stopping. Figure 11 shows three positions where the detection device II stopped.

Fig. 11
figure 11

Three ending states of detection device: a halfway stopping; b accurate arrival; and c cross

Discussion

Selection of fitting methods

The R2 value of linear fitting in Fig. 10 is greatly affected by the selected time period. If the ratio of the R2 value of linear fitting to the R2 value of exponential fitting is δ, the relationship between δ and the location of the selected time period may be determined. The data is divided into 21 segments with each time length 5 s. The data of the odd number segment is removed in order as well as the data of the 0–5 s period (the measured value of some sensors is zero). Linear and exponential fitting is performed on the mean value of each sensor in each section of data and the height of the sensor. R12 represents the correlation coefficient of linear fit and R22 the correlation coefficient of the exponential fit. The results of R12 and R22 over time are shown in Fig. 12.

Fig. 12
figure 12

Relationship between R12, R22, δ and selected time period

In Fig. 12, R12 is greater than 0.98 and the data points are stable. R22 is close to R12 in the 70–75 and 100–105 s data, but fluctuates greatly in other periods. The δ value shows the difference between R12 and R22. It concludes that the exponential decline law is more suitable than the linear decline law to describe the change of CO concentrations directly above the ground of β-type residual fires, and is not affected by the location of the selected time period.

Reliability

In the 25 experiments in Table 3, detection device II does not illustrate misjudgment but shows that the detection system is reliable. To further verify this reliability, a validation test was carried out whereby the eight CO sensors were divided into two groups, four in each group. Each group of CO sensors was connected in series with hexagonal copper columns to ensure a spacing of 200 mm (the length of 8 hexagonal copper columns). Two new screws with a thread diameter of 3 mm were placed on the long board to ensure that they were 300 mm from the first set of screws in the length direction of the board. Two sets of CO sensors were connected to two sets of screws so that the first set was directly above the combustion basin. The sensors, development boards and computers were connected with Dupont wires and the smoldering charcoal was placed into the combustion basin for measurement. Each sensor collected data for 105 s simultaneously. The average of the data of each sensor for 50–55 s were used to draw Fig. 13, with the sensor height as the independent variable.

Fig. 13
figure 13

Characteristics of mean CO concentration at the same height above and side of smoldering wood carbon

In Fig. 13, the first set of data shows the characteristics of exponential decline, and the R2 value is higher than 0.98. The difference between the two sets of data is mainly in two aspects, one is the first group of decreasing data and the second group is increasing data. The first group of data is greater than the amount of data in the second group. These differences ensure the accuracy of the results and verify the conclusions. After removing the data of the first 5 s, five segments of data were randomly selected for verification, and the results were the same as those of 50–55 s data.

Improvement of search algorithm

Reducing the detection time can improve efficiency while still ensuring accuracy. In the search algorithm, each step of judgment consumes 5 s, so reducing the total number of program judgments can significantly reduce the search time. We introduced a step factor ε into the search algorithm so that its size was positively correlated with the height of the detection device II (Table 4). The initial state of the detection device II was 850 mm above the ground, and its step size was 20 cm in the traverse state. When the detection device II detects CO gas, the value of ε is according to Table 4. Five experiments of group 6 were carried out under the same experimental conditions. The results show that the five new experiments were successful and the average time was 55 s, 39 s less than that of group 3 (h = 5 cm), even 3 s less than that of group 4 (h = 10 cm).

Table 4 Value rule of ε

The introduction of the step factor greatly reduces the time consumption, but the average value is still high. In the above six groups of experiments, the forward direction of the detection device II was always unchanged, resulting in low forward movement efficiency. Adjusting the detection device II to the windward direction will improve the forward movement efficiency and reduce the time consumption. The rotation factor θ is Eq. 1:

$$\uptheta = {\text{K}} \cdot \left| {\frac{{{\text{P}}1 - {\text{P}}3}}{{max\left( {{\text{P}}1, \,{\text{P}}3} \right)}}} \right|$$
(1)

where K is the base of angle compensation; P1 is the instantaneous measurement value of sensor 1 of detection device II, and P3 is the instantaneous measurement value of sensor 3; max (P1, P3) is the larger of the two measurements.

The search program is designed to rotate θ in the direction of larger values of sensor 1 and sensor 3 before each detection device II moves forward. Within the range (0°, 90°), different K values were taken at 10° intervals. Experiments were conducted under the same conditions. Figure 14 shows the search state of the detection device II after adding the rotation factor with the results as a line chart (S–K) of how the search time changes with the K value (Fig. 15). It can be seen from the figure that the smoldering fire was successfully detected in eight experiments, and the search time of each experiment was less than 55 s. When K was 40°, the search time was only 31 s.

Fig. 14
figure 14

Detection after improvement of search algorithm: a detection device ready to rotate; b smoldering wood carbon detected

Fig. 15
figure 15

Search time corresponding to different K values

In order to get a reasonable value of K, the smoldering charcoal was 5, 10, 15 and 30 cm long under the same conditions, and the same experiment was carried out (Fig. 16).

Fig. 16
figure 16

S–K line chart of smoldering wood carbon of different sizes

The results show that with the change of K from 10° to 80°, the five lines all show the pattern of first decreasing and then increasing. When the size of the smoldering charcoal was 5 and 10 cm, the corresponding two S-K lines have a minimum value when K was 30°. In the other three groups of experiments, the corresponding three S-K lines are all a minimum value when K was 40°. This indicates that the optimal value of K is affected by the size of smoldering charcoal. The role of the rotation factor θ is to improve the efficiency of the forward movement of the detection device II, and the selection of K should conform to this principle. In the experiment, excessive K values caused the detection device to spiral down at the end of the search process, increasing more time consumption. When the K value was in the 30°–40° range, good results can be obtained.

Conclusion

To better detect hidden smoldering fires below the surface of burned forest lands, we took CO gas detection as a breakthrough to study the diffusion law of the gas in the near-surface space of this kind of residual fire. It was found that the average CO concentration in the center of the gas mass decreased exponentially with the increase of height in a windless state. Taking this as the judgment condition, it was verified that detection device II composed of eight CO sensors could detect β-type residual fires under weak wind condition. The rotation factor θ and step factor ε were added to optimize the search process. The residual fire detection robot equipped with the detection device II greatly improved detection accuracy and reduced the probability of residual fire reignition. In this study, the use of smoldering charcoal instead of β-type postfire was one dimensional and research on smoldering humus layers should also be included in future. The actual fire area is also large and the shape complex. The path planning problem of the residual fire detection robot needs to be further studied. In addition, the reliability of detection device II is greatly affected by the wind speed, and the detection results only have certain reference value under strong wind conditions.